EXTERNAL APPROXIMATIONS BY FINITE ELEMENTS
Victor Apanovitch, SIMSOLID Corporation
Proprietary & Restricted Rights Notice
© 2016 SIMSOLID Corporation All Rights Reserved. This document is an unpublished draft of book “External Approximations by Finite Elements” by Victor Apanovitch. It is intended to provide background information on the mathematical foundations used in the commericial software application, SIMSOLID. Unauthorized use, distribution or duplication is prohibited.
Table of Conte ont ents nt s Preface Introduction 1. Partitions, spaces, and boundary operators 1.1. Basic definitions 1.2. Spaces associated with a partition 1.3. Generalized Generalized Green’s formula 1.4. Properties of the boundary operators
2.
The hallmarks of external approximations by finite elements 2.1. Approximations on fixed and variable partitions
2.2. More spaces associated with a partition 2.3. Some auxiliary assertions 2.4. The hallmarks of external approximations
3. Types of finite elements for external approximations and their construction 3.1. Finite elements conformed by a subspace 3.1.1. Element definition 3.1.2. Construction of basic functions of an element 3.1.3. A sufficient condition for existence of an element basis 3.1.4. Virtual finite element 3.2. Homogeneous and composite finite elements 3.3. Spaces of finite elements
4. Error estimates and convergence of external approximations of Sobolev spaces 4.1. Operators of approximations 4.2. Convergence of approximations on a fixed partition 4.3. Convergence of approximations on a family of partitions 4.3.1. Some definitions 4.3.2. Error estimates for piecewise polynomial approximations 4.3.3. Convergence of approximations
5. Approximations of boundary spaces 5.1. Definitions 5.2. Operators of boundary approximation 5.3. Approximations on a fixed partition 5.4. Approximations on a family of partitions
6. External approximations of variational equations 6.1. Construction and total error estimate 6.2. External error estimates and convergence of approx imations 6.3. Convergence of approximations in case when bilinear form is not V -elliptic
7. Examples of finite elements 7.1. Justification of some classic non-conformed finite elements 7.1.1. Wilson's element
7.1.2. Family of Crouzeix-Raviart ‘s elements 7.1.3. Morley's element 7.1.4. Fraeijs de Veubeke’s element 7.2. Elements for external approximations 7.2.1. 2D elements for boundary value problems of second and fourth orders 7.2.2. 3D elements 7.2.3. Elements for approximations of spaces of vector functions 7.2.4. Axisymmetrical elements 7.2.5. Finite elements for stress concentration
8. Numerical experiments 8.1. Torsion 8.1.1. Formulation 8.1.2. Model problem 8.1.3. Stress concentration at a semi-round notch 8.2. Thin plates bending 8.2.1. Formulation 8.2.2. Bending of a cantilever plate 8.2.3. Bending of a square plate 8.3. 2D elasticity problems 8.3.1. Formulation 8.3.2. Extension of a square plate 8.3.3. Bending of a circular bar
8.3.4. Extension of a plate with a round hole 8.4. Special elements in stress concentration 8.4.1. Extension of a plate with an elliptic hole 8.4.2. Extension of a strip weakened by a square hole with filleted corners 8.4.3. Extension of a perforated plate 8.4.4. Strip with a round hole 8.4.5. Extension of a strip with a cutout 8.5. 2D problems of linear fracture mechanics 8.5.1. Plate with an inclined crack 8.5.2. Strip with a crack 8.6. Axis-symmetrical elasticity problems 8.6.1. Formulation 8.6.2. A hollow sphere under internal pressure 8.6.3. Stress concentration around a spherical cavity 8.6.4. Round plate under pressure 8.6.5. Spherical dome under pressure 8.6.6. Sine-shaped shell of infinite length under pressure 8.6.7. Shell compensator Reference list
Preface The Finite Element Method proved to be a powerful analysis tool in many fields of engineering. Its wide use and still great potential stimulate research and development of new effective finite element based technologies. This book presents an innovative approach to the finite element approximations. The approach possesses some new unique features and at the same time it inherits the computational advantages of the conventional finite element analysis. For example: finite elements of arbitrary shape with approximating functions of arbitrary type can be easily formulated; the types of the approximating functions do not depend on the element’s geometry; the element is not mapped onto a canonical shape; the matrix of the equation system is sparse, symmetric, and positive definite; and others. I believe that the suggested concept will lead to new analysis technologies created on its basis.
Generality and constructiveness are the major features of the theory proposed. All the results are obtained in abstract form under minimal restrictions on finite element geometry and contents of approximating spaces. It provides wide applicability of the suggested theory to different areas of analysis and simulation. It is described in detail how the methods of construction of the finite elements and approximations follow from the basic theorems, and how they are implemented in numerical algorithms. Numerous benchmark problems illustrate the applicability and accuracy of the method.
Section 1 introduces the definitions of spaces and operators associated with a partition of the domain into finite elements. Section 2 presents the necessary and sufficient sign, as well as two sufficient signs of the external finite element approximations of Sobolev space H m () ( m 0 is
arbitrary integer) under minimal limitations on the shape of finite elements (it is assumed that the boundary of the finite element K must be sectionally smooth only) and the contents of spaces of approximating functions (it is assumed that the functions just belong to the space H m ( K ) ). Herewith two types of the external approximations are considered: one is associated with a fixed partition of the domain into finite elements, the other is associated with a family of partitions. Section 3 addresses the formulations of different classes of finite elements for external approximations. Section 4 introduces the operators of approximations and studies their convergence properties on fixed and variable partitions of the domain. Section 5 develops the approximations of spaces of functions determined on the boundaries of the partition. These spaces are used to construct external approximations by necessary and sufficient sign. Section 6 studies the external approximations of variational equations. Section 7 justifies some classic nonconformed finite elements as well as presents examples of finite elements for external approximations. Section 8 discusses results of numerical experiments on numerous benchmark problems.
This book should appeal to numerical analysts and software developers in different fields of engineering such as structural and dynamic analysis, applied mechanics, multiphysics, and computational mathematics.
The theory presented in the book was developed in 1981-1991 when I worked as a professor of Department of Theoretical Mechanics at Belarus State Polytechnic Academy. I take this opportunity to thank my former colleges for the support.
Victor Apanovitch
Introduction Elliptic boundary value problems can be stated in two equivalent forms: differential formulation or weak formulation. The equivalency means that under the assumption of sufficient smoothness of the input data (for example, the smoothness of the coefficients of the differential operator of the problem and the smoothness of the boundary of the domain) both formulations can be transformed one to another and their solutions coincide. The idea of a weak solution of a boundary value problem generalizes the idea of its classic solution.
In the weak form a boundary value problem is stated as follows: find element u V that satisfies an abstract variational equation
v V
a u, v
f v
,
(1)
where:
V is an enclosed subspace of Sobolev space (or some product of such subspaces)
a u, v is a continuous on V V bilinear form
f v is some linear form on V
In computations, the infinite-dimensional space V is approximated by a finite-dimensional space X h . Herewith the continuous variational problem (1) is approximated by the discrete one as
follows: find element uh
X h that satisfies the discrete variational equation
vh X h
a uh , vh
f vh .
(2)
In result, the equation (1) is reduced to a system of linear algebraic equations from which the approximant uh of the exact solution u of the variational problem (1) is found.
In Finite Element Method (FEM) the space X h is constructed by partitioning the domain into non-intersecting sub-domains (finite elements (FE)), and by defining on each FE smooth approximating functions (shape functions) that satisfy specific compatibility conditions on interelement boundaries. The fulfillment of the compatibility conditions is the major problem for the construction of FE approximations of the space V . These conditions define the smoothness requirements imposed onto the approximating functions from the space X h of finite elements. In conventional FEM, the space X h is constructed in such a way that the following inclusion holds X h
V
(3)
If the spaces X h meet the condition (3) then the corresponding approximations of the space V and equation (1) are called internal approximations. For the internal approximations the closeness of the approximate and the exact solutions is evaluated by the norm of the space V . Herewith the unique solvability of the continuous variational problem (1) guarantees the unique solvability of the discreet variational problem (2), and the density of the family X h in the space V guarantees the convergence of the approximate solutions to the exact one. Finite elements that
generate internal approximations are referred to as conformed FE.
The construction of the conformed FE of high accuracy to approximate the Sobolev space m
H
is
not a simple task. When
m 1 then the functions from the space X h must be
continuous, but their first derivatives can be discontinuous on inter-element boundaries. When
m 2 , the approximating functions must be continuous with their first derivatives. It
necessitates the use the complex structures of degrees of freedom even for finite elements of the simplest form (triangle or quadrilateral for 2D problems). The problem becomes more complicated in case of curved boundaries. Such boundaries have to be simulated with elements that have non-planar faces. The curved elements are constructed through the mapping of some initial FE of simple form. However, such mapping worsens the approximating properties of the FE, particularly, if the FE is significantly distorted.
It is much simpler to construct FE which do not meet the condition (3). Such FE are called non-conformed FE. However, the justification of non-conformed FE is a complex mathematical problem. To solve it, one has to accomplish the following steps:
introduce suitable spaces to compare the approximate uh and the exact u solutions (the space V can not be used for that because X h
V )
prove the existence and the uniqueness of the solution of the discreet variational problem (2) (these features are not provided automatically like in the case of internal approximations)
estimate the non-conformity errors and prove the convergence of the approximate solutions to the exact one
Due to the mathematical difficulties, the justification of non-conformed FE often is not conducted, and the accuracy and robustness of the approximations are verified through numerical experiments on benchmark problems only. Nevertheless, the efficiency of the non-conformed schemes caused their wide use in computations.
Problems of the mathematical justification of non-conformed FE were considered in works [1 28]. There had been developed several justification techniques. In 1965, in work [1] the patch test was introduced as the criterion of convergence of non-conformed FE. The patch test has been used to justify many non-conformed approximations applied to the specific problems [2 8]. In work [9] a variational treatment of the patch test was developed and a general estimate of the non-conformity error of the approximations of the variational equations was derived. The non-conformity error accounts the discontinuity of derivatives up to
m 1 -order
of the
approximating functions from the space of FE at the inter-element boundaries. The proof of convergence of a non-conformed FE usually consists in obtaining the estimates of the non-conformity errors [5, 13, 14, 18 – 20, 29, 30]. In works [9, 12, 31] there was demonstrated that for a particular type of problems, the success of patch testing is a sufficient condition for the convergence of some non-conformed FE. Some of the non-conformed schemes were analyzed in work [17]. The results of [17] were generalized in work [24] by establishing some sufficient conditions of convergence. The iterative method of solution for schemes with non-conformed elements was suggested in work [27].
In 1980, in work [23], there was developed an approximation that successfully passed the patch test, but did not converge. Thus, it was demonstrated that, in general, the patch test is not a sufficient condition for the convergence of non-conformed FE. The generalization of the patch test was given in work [22], in which have been established the necessary and sufficient conditions of convergence of the non-conformed approximations. The generalized patch test did not depend on the type of the boundary value problem, but it was not constructive because in fact
it just stated that the error of non-conformity of the approximation must tend to zero when the mesh of FE is refined. So, the justification of the specific non-conformed FE was performed in work [22] for the model boundary value problems using the conventional technique for obtaining the estimates of the non-conformity errors.
In the above-mentioned works, the justification of the specific non-conformed FE was performed for specific types of boundary value problems, and the properties of the corresponding equations were extensively utilized. In such justifications, the proof technique usually assumed that a subspace of conformed FE was present in the space of non-conformed FE. This assumption reduced the applicability of the results obtained and prevented the development of general approaches to the construction of non-conformed approximations. In fact, the suggested criteria only helped to find out what kind of non-conformed functions can be added to the basic functions of conformed FE in order to improve accuracy on rough partitions of the domain.
Another approach to treat non-conformed approximations is based on the concept of the external approximations introduced in work [32]. The general theory of the external approximations of Sobolev spaces and variational equations has been developed in work [33] where were also given the examples of its application to the finite difference approximations and the approximations by parts. In terms of the theory of external approximations some non-conformed FE have been studied in works [11, 28, 34].
Let us note that the problem of justification of non-conformed FE is closely related to the problem of combination of approximations of different types on sub-domains, and with the
problem of approximate fulfillment of the essential boundary conditions. Some methods of the combinations have been studied in works [35, 36]. In work [37] it has been demonstrated that the non-conformed basic functions can lead to a significant error in the solution. The combined approximations are important when solving the boundary value problems in presence of singularity in the solution, and for unbounded domains. For example, in works [38, 39] the non-conformed combined approximation has been successfully applied for the solution of a boundary value problem for a domain with a cut. Herewith in the circle that contained the tip of the cut, the asymptotic representation of the solution for an infinite domain with a cut has been used. On the rest of the domain the solution has instead been approximated using linear triangular FE. In work [40], in terms of the theory of external approximations, it has been studied the approximate fulfillment of the essential boundary conditions in a thin plate bending problem. In work [41] it was shown that the methods of arbitrary accuracy could be developed through the replacement of the interpolation of the boundary conditions with the orthogonalization of the boundary error to some system of boundary functions.
The external approximations of Sobolev spaces and variational equations by finite elements have been systematically studied by the author in works [96-100]. This book generalizes the results of these works.
1. Partitions, spaces, and boundary operators Basic definitions Let us be an open bounded domain in R n . For any integer m 0 , the Sobolev space H m consists of functions v L2 for which all generalized derivatives D v, m (where
1 ,..., n is a multi-integer, and a1 ... n ) belong to the space L2 of all functions
integrable with square, i.e.
, m
D v dx
D
1
D v L2
vD dx
where dx dx1...dxn ; D is the space of all infinitely differentiable functions which have compact supports in .
In the space H m , norm and semi-norm are defined as follows:
v
m,
m 2 D v dx 0
2 v k,m D v dx k 1 k m
1
2
1
2
The Sobolev space H 0m is defined as a closure of the space D by the norm m, .
A bounded domain in the Euclidean space R n is said to belong to the class C k , where k is a non-negative integer, if there exists the finite number of local coordinate systems and local mappings f r , 1 r R , as well as real numbers 0, 0 such that the following conditions are satisfied (see Fig. 1.1):
functions f r , 1 r R are k -times continuously differentiable in a closed (n-1)-cube
x
r
ˆ
Rn 1 : xnr
Here
x1r , ..., xnr 1 , x r
x r ˆ
ˆ
means that xir
, 1 i n 1
the boundary of the domain is R
x , x ; r
ˆ
r n
xnr
r 1
x r , x r ˆ
ˆ
in any local coordinate system 1 r R
x , x ; x , x ; r
ˆ
r
ˆ
r n
fr
r n
fr
x x f x , x x f x , r
ˆ
r
ˆ
r n
r
r n
r
ˆ
r
ˆ
r
xr ˆ
x ˆ
r
It is said that any domain of class C 0 has a regular (or Lipshitz) boundary. The boundary (or its part S ) is called smooth if f r , 1 r R belong to the class C (i.e. if the boundary is a (n-1)-dimensional infinitely differentiable manifold in R n ). The boundary of the domain is called sectionally smooth if it can be divided into finite number of pieces i , of class C , so that
N
i . i 1
Spaces associated with a partition Let us be a bounded domain in R n with a sectionally smooth boundary . Let us denote a partition of into a finite number of sub-domains K (finite elements). Let us the partition fulfills the following conditions:
K K
for any K , domain K is closed and the set of its inner points K is not empty (i.e. K
0 )
for any pair of elements K 1 , K 2 : K
1
K 2 0
for any K the boundary K is sectionally smooth and regular
Partitioning the domain into FE, let us define the space V
H K
m
K ,
K
.
(1.1)
Mutually orthogonal subspaces
0, 0, ..., H
K , ..., 0 are contained in the space
m
V . These
subspaces can be identified with the spaces H m K . Thus, the space V is a direct sum of its mutually orthogonal subspaces V
H
m
K ,
K
,
K
and any element v V can be represented as follows v
v
K
,
K ,
(1.2)
K
where element 0, 0,..., v K ,..., 0 is identified with an element v K H m K .
The inner product and the norm in the space V are defined as follows:
u , v V u K , v K m, K
K m K K D u , D v 0, K K 0 m K K D u D v dx 0 K K 1 2 2 K u K m , K 1 2 2 m D u K dx K 0 1 2 2 m K D u dx 0 K
u
V
One can see that the norm
m,
is the restriction of the norm V on the space H m .
(1.3)
To approximate the space H m let us define a finite-dimensional space, also referred to as a space of finite elements X h
P
K h
V,
K
K
where PhK are some finite-dimensional spaces of functions determined on the sub-domains K .
For any FE K the space PhK approximates the space H m K ; the finite element space X h approximates the space V . Approximating functions from the space X h are sufficiently smooth on each FE of the partition ( PhK H m ( K ) ) but they are discontinuous across the inter-element boundaries. Thus the following inclusions take place X h
V L2 ,
and approximating functions from the space of FE do not belong to the space H m . The functions are “external” with respect to the space H m and the concept of external approximations must be used to justify them.
The formal definition of external approximations is as follows (see [33]). Assuming that:
h is a family parameter
V h is a finite-dimensional space with some norm
ph is a continuous linear mapping from V h onto V (an extension operator)
r h is a mapping from V onto V h (a restriction operator)
approximations Vh , ph , r h of a space V are called the external approximations of the closed subspace V from V , if the family Vh , ph , r h possesses the following property: if phuh u weakly in V then u V . Approximations Vh , ph , r h of the space V are called converging in V if
u V
u ph rhu
V
0.
Let us assume that V H m , and the space V , associated with the partition , is defined by (1.1). Then V V is a closed subspace.
Let us also define a finite-dimensional space of vectors Vh
R N h .
Extension operators ph : uh
Vh ph uh X h
(1.4)
are defined by the inclusion X h V .
The primal goal is to establish conditions that must be met by the approximating functions ph uh
X h L2 () in order to restore the smoothness properties of the functions in limit, when
the weak convergence phuh u in the space V implies the inclusion u H m .
Generalized Green’s formula
The validity of generalized Green’s formula is proved in [33] under the following assumptions. Let us there be given Hilbert spaces V and H , and the operator L V , T , such that:
maps V onto T
H contains V , and V has a stronger topology
the kernel V 0 of the operator is dense in H
Further, let us there be given a continuous on V V bilinear form a u , v and a linear on V 0 functional u defined by the relation
u , v a u , v
v V0 .
The formal operator is a continuous linear operator from V into V 0 . The domain of determination of the operator is a Hilbert space V
u V
u H
with the norm u V
u
2 V
u
2 H
1
2
.
Under the given assumptions there exists a unique operator such that the following Green’s formula is valid (see [33])
u V , v V
a u, v
u , v u , v ,
(1.5)
where , is the inner product defined on H H ; , is the duality pairing defined on T ' T (here T is a conjugate space)
Various Green’s formulas can be established from the general formula (1.5) by choosing of bilinear forms and spaces. Such formulas are widely used in the theory of e lliptic boundary value problems and in particular when proving the equivalence of weak and differential formulations of the problems. To determine the hallmarks of the external FE approximations of Sobolev spaces let us introduce specific spaces, operators, and generalized Green’s formula associated with the partition of the domain.
Let us partition the domain into FE, then let us determine the space V by the formula (1.1), and let us introduce the continuous on V V bilinear form a u , v
a u
K
K
, vK
K
where K
K
u
K
a u , v
K
u
K
K
D v dx
K
u
K
,
v
v K ,
K
m
The formal operator , associated with the form a u , v , is defined by the expression
u
K K u
,
K
K
where
K u K 1 D u K ,
m
The domain of determination of the operator is the space V
H m K V . K
Let us define a face K r of the FE K as a (n-1)-dimensional sub-domain of the boundary K , which either divides the adjacent elements or coincides with a patch of the boundary of the domain (Fig. 1.2). Let us introduce the following notations:
K r is the set of all faces of the FE
K
M
M K r ,l is the set of all smooth face sections K r , l of the face
K r (in brackets a typical
element of the set is indicated) In these notations symbols r and l identify the face and the smooth face section respectively (they are not indices).
Now the following spaces are introduced:
H
L K L H K
2
2
K
V0
m
0
K
T
m 1
H
m j 1 2
Kr Kr ,l j 0
K
K r , l
where K , Kr M Kr , Kr , l M Kr, l .
The space T contains mutually orthogonal subspaces
0, 0, ..., H
m j 1 2
K r ,l , ..., 0
that can be identified with spaces H m j 1 2 K r , l . Then
T
m 1
H K
K r Kr ,l j 0
m j 1 2
K r ,l
(1.6)
and any element g T can be represented as follows
g
m 1
g K
K r ,l , j
.
K r Kr ,l j 0
The trace operator associated with the partition is defined as follows
m 1
K r ,l , j
K
,
(1.7)
K r Kr ,l j 0
where rK , l , j is an operator of j-order differentiation along the outward normal to the smooth section K r , l of the face K r of FE K . When j 0 , the function rK , l , 0 u is a restriction of the function u on K r , l .
According to the trace theorem from [33] each component rK ,l , j of the operator is a continuous linear mapping from the space H m K onto the space H m j 1 2 K r , l . Hence the operator maps the space V onto the space T . Since the space V is enclosed into the space H with stronger topology, and the space V 0 is a kernel of the operator that is dense in the space H , then the assumptions of the Theorem 6.2.1 from [33] are met. So, for any , m there
exists a unique operator such that for any u , v V the following generalized Green’s formula is valid a u , v
where:
u , v u , v ,
(1.8)
m
Kr Kr ,l j 2 m 1
K
rK, l,, j m
,
rK , l , j
L H m K , H m j 1 2
j 2m 1
K r , l
Taking into account the representation (1.2), the formula (1.8) can be rewritten as:
u , v V
u
K
D v dx 1
K
D u
K
K
vdx u , u ,
(1.9)
where: uK
u K ; v K v K .
The operators and are referred to as boundary operators.
Properties of the boundary operators The Sobolev space H m consists of functions v L2 such that all their generalized derivatives D v, m belong to the space L2 , i.e. for any multi-integer , m there exists such a function D v L2 that
D
D v dx
1
vD dx 0 ,
(1.10)
where D is the space of all infinitely differentiable functions that have a compact support in
.
Since the space D is dense in the space H 0m , it is possible to extend the expression (1.10) onto the space H 0m , i.e. it holds for any function H 0m .
Let us v denotes the function that is equal to the function v V in and that equals zero in a ˆ
compliment of in R n . It is known that if v H m R n then v H 0m . Thus, if for any ˆ
function D R n from the space of infinitely differentiable functions, the expression below holds
D v
R n
ˆ
K
K
dx 1
D vdx 0 ,
ˆ
(1.11)
Rn
then v H m R n and v H 0m . Here v K v K . ˆ
ˆ
ˆ
Because of the density of the space D R n in the space H m R n , the expression (1.11) can be
extended by continuity onto the space H m R n . Integrals in (1.11) are evaluated only on the domain because v 0 outside . It must be also noted that, since by assumption the domain ˆ
has a regular boundary, the space H m coincides with the space of restrictions on of all
functions u from the space H m R n .
Now let us consider the space V defined by (1.1). Spaces H m and H 0m are the closed subspaces of V . The conditions to determine if the function v V belongs to the space H m or H 0m readily follow from the formula (1.9). In fact, if the following inclusion holds
u
u
K
H m
K
then D u
D u
K
.
K
Thus, using (1.9)-(1.11), the following lemma can be formulated.
Lemma 1.1 A function v V belongs to the space H m -and respectively to H 0m - if, and only if, for any u H 0m -and respectively for any u H m - and for all , m , one of the two equivalent conditions holds:
u D v
K
u, v
K
dx 1 D uvdx 0
0
Now, some properties of the boundary operators and that act in the spaces H m and m
H 0
can be established. Let us introduce the following designations:
r is the common face of adjacent FE
M r is the set of common faces of all elements of the partition
M r ,l is the set of common smooth face sections of all elements of the partition
K and K 2 1
For any so that
r ,l M r ,l there exist the face sections K r1,l and K r2,l of adjacent FE
K and K 2 1
r ,l Kr1,l K r2,l . The j-order derivative of a function u K H m K along the outward
normal to the face section K r ,l is defined as follows rK, l , j u K
! D u j !
K r,l
n
j
where
1,..., n is a multi-integer
nr ,l
! 1 !2 !...n !;
n1 ,..., nn is an outward unit normal (see Fig. 1.2.b) vr ,l
v1 v2 ...vn 1
2
n
1 2 The outward normals to the common face section of the adjacent elements K and K are
opposite. Hence for any function v H m the following relation holds
r , l M r , l
1r , l , j v1
j
1 r2, l , j v 2 ,
0 j m1
where
v1
vK
1
,
v2
v K
2
.
Here and below superscripts 1 and 2 denote the functions and the operators associated with the 1 2 adjacent elements K and K .
Therefore, the operator maps the space H m onto the closed subspace T of the space T characterized as follows
T g ˆ
1
gr , l , j
m 1
g
K r , l, j
T
K r Kr ,l j 0
K
j
1 g r2, l , j ,
r , l M r, l
(1.12)
0 j m 1
Identically, considering the space H 0m it is apparent that the operator maps it onto the closed subspace T0 ˆ
g T ˆ
g
0 .
(1.13)
Now some properties of the boundary operator can be established.
Lemma 1.2
Operator ,
m maps
the space H 0m onto the orthogonal complement T of ˆ
' the space T in the space T ˆ
Operator ,
m maps
the space H m onto the orthogonal complement T 0 of ˆ
the space T 0 in the space T ˆ
◊ Proof – The
operator maps the space H m onto T . According to Lemma 1.1:
, m u H 0m , v H m
u , v
0
u H m , v H 0m
u , v
0
Now the characterizations of the spaces T and T 0 can be established.It has been demonstrated ˆ
ˆ
that:
f T0 , g T0
f, g
0;
(1.14)
f T , g T
f, g
0.
(1.15)
ˆ
ˆ
ˆ
ˆ
Taking into account the definition (1.13), the expression (1.14) can be rewritten as follows f , g
m 1
K
K r Kr , l j 0
m 1
r ,l j 0
m 1
r ,l j 0
K
K
fr ,l, j , g r,l , j
f r2, l , j , g r2, l, j
f r1, l , j , g 1r , l , j f r1, l , j
(1.16)
j
1 f r2, l , j , g 1r , l , j 0
, K r M K r , K r , l M K r , l , r, l M r, l .
where K
Since (1.16) must hold for any g T , then the space T 0 is defined by the characterization ˆ
ˆ
T0 ˆ
f T ' | r , l M r , l
1
f r, l, j
f
1 f r2, l , j ,
0 j m 1 .
Identically from (1.15) one can obtain T ˆ
f T0 | Kr, l M Kr, l ˆ
f r, l , j
0,
0 j m 1
where M K r, l is a set of smooth face sections of FE that coincide with the boundary domain .
of the
Figure legends for Section 1:
Fig. 1.1 To the definition of the domain boundary
Fig. 1.2 Adjacent finite elements: a) linked along the common boundary b) separated
2. The hallmarks of external approximations by finite elements Approximations on fixed and variable partitions Two types of the approximations are under consideration:
The first type is associated with a fixed partition of the domain . This type is an analogy of p-version of FEM
The second type is associated with a family h of partitions. This type is an analogy of hversion of FEM
Let us define a fixed partition of the domain and let us associate with it a family X h of FE spaces X h
PhK ,
K
K
K
where Ph
H m K are finite-dimensional spaces.
The absence of the parameter h in the partition designation indicates that each space X h is associated with the same partition . The dependence between h and approximations must be such that: when h 0
dim PhK .
For instance, if PhK is a space generated by N basic functions, then h 1 N .
Second type of the approximations is associated with a family h of partitions. In this case the parameter h has the meaning of maximum diameter of FE: h
max hK , K
hK
diam K .
It is assumed that the family h is regular so:
there exists such positive constant that
K h
hK K
(2.1)
where:
hK
diam K ,
K
sup diam S , S is a sphere from K
diameters hK approximate zero
The family h is constructed through the partition refinement. Spaces PK of the approximating functions are fixed; the number of FE - card h - tends to infinity when h tends to zero. The FE space is defined as follows X h
P K ,
K h .
K
For the approximations associated with the fixed partition , definitions (1.1) and (1.6) of spaces V , V0 , H , T contain finite products of spaces. For approximations associated with the family h of partitions, the finite products in expressions (1.1) and (1.6) are substituted by numerable
products, because when h 0 then card h
.
In this case, for instance, the space V
consists of infinite vectors u
u1 , u 2 , ..., u K , ... ,
u
K
H m K
such that u
2 V
u
K
K
2 m , K
.
More spaces associated with a partition
In order to establish the signs of the external FE approximations of Sobolev spaces, the duality pairing
,
on T ' T in formula (1.8) is to be substituted by the inner product in the
corresponding space. Let us introduce the space m 1
HT
L2 K r , l (2.2) K
K r Kr ,l j 0
with the norm
g
m 1
g K
Kr Kr ,l j 0
K r, l, j
HT
g
H T
m 1 g rK , l, j K K K j 0 r
r ,l
1
2
0, K r ,l
2 .
Space T , defined by (1.6), is continuously and densely enclosed into the space H T . According to the Theorem 2.1.5 from [33], T and H T can be identified with subspaces which are dense in the conjugate space T ' : T the duality pairing
,
HT T ' , and these inclusions are continuous and dense. Then
on T ' T can be identified with the unique extension of the inner
product , H in H T . Further in the text the closures of the spaces T , T 0 in the space H T are ˆ
T
ˆ
designated as H and H 0 , and the orthogonal complements of these closures in H T as H and ˆ
ˆ
ˆ
H 0 correspondingly. Thus: ˆ
f H T | r , l M r , l
H 0 ˆ
H ˆ
f H 0 | ˆ
1
f r, l, j
f
j
1 f r2, l, j , 0 j m 1 (2.3)
0
(2.4)
Functions that are determined on the face K r can be represented through their components determined on the face sections as follows f rK, j
f
K r ,l
K r ,l , j
L2 K r ,l ,
0 j m 1.
Then the characterization (2.3) can be rewritten as follows
H 0 ˆ
f HT | r M r
f r1, j
j
1 f r2, j , 0 j m 1
(2.5)
Characterizations of the spaces H and H 0 can be rewritten in the same manner. ˆ
ˆ
It must be mentioned that, for the fixed partition of the domain, the space H T is a fixed separable Hilbert space. Therefore, there exists a family of closed subspaces that is dense in H T . When the family h of partitions is under consideration, then the family of spaces H Th , defined in accordance with (2.2), take place. When h 0 the product of finite number of spaces in (2.2) becomes a product of numerable spaces. In this case a family of spaces Rh is called dense in space H T if, for any sequence xh
H Th , there exists such a sequence yh Rh that lim xh yh h 0
H Th
0.
Some auxiliary assertions Now let us consider two auxiliary lemmas.
Lemma 2.1 Let us M be a Hilbert space and N M is a closed subspace. A bounded sequence h
M
satisfies the relationship
N
lim , h 0
(2.6)
h0
if, and only if, there exists the family of closed subspaces Gh
N which is dense in the space
N and, for any h , it possesses the following property
gh Gh
gh , h 0
Necessity - The space M can be represented as a direct sum of the subspaces M
(2.7)
N N ,
where N is an orthogonal complement of the subspace N in the space M . Then any element of the sequence h can be expressed uniquely as follows h
1h h2
1h
N ,
h2 N .
For any h , let us F h1 be a one-dimensional subspace generated by the element 1h , i.e.
f Fh1
f
1h ,
R .
Because the space F h1 is one-dimensional, therefore the orthogonal complement Gh of the space 1 F h in the space N does exist and has a co-dimension equal to one.
The space N can be represented as a direct sum of its mutually orthogonal subspaces, i.e. N
Gh F h1 . Hence any element N can be uniquely represented as follows
h h h1
h
Gh . R, h
Then 2
lim , h lim h h h1 , 1h h2 lim h h1 0 h 0
h 0
h 0
and h 1h 0 strongly in M . That implies
lim h hh1 lim h . h 0
h 0
Thus Gh is the family of the subspaces that is dense i n N and possesses the property (2.8).
N Th en for any element Sufficiency - Let us the relationship (2.7) holds. Then lim , h lim , h gh , h lim gh , h lim h , h 0 . h 0
h 0
h0
h0
Here h is a projection of onto the closed subspace Gh
N .
Lemma 2.2 Let us M , N and N and h fulfill the assumptions of Lemma 2.1. If there exists a sequence g h
M
such that
N
, gh 0 ,
lim h g h 0 , h 0
then (2.6) holds.
It ◊ Proof – It
follows from (2.8) that
N
lim , h lim , h gh . h0
h 0
(2.8) (2.9)
Then (2.6) follows from (2.9) and the inequ ality
, h g h
h g h .
The hallmarks of external approximations approximations The necessary and sufficient sign of external FE approximations of a Sobolev space is defined by the following theorem.
Theorem 2.1 The FE approximations
Vh , ph , r h are
external approximations of the space H m -and
space H 0m respectively- if, and only if, the following c ondition is met
g h Gh , ph uh X h
g h , p huh H 0 ,
(2.10)
T
where Gh is a dense family of closed subspaces of the spaces H -and H 0 respectively. ˆ
ˆ
m
Proof – Let us consider the approximations of the space H
.
Let us assume that a
sequence phu h weakly converges in the space V . It means that, for any function H
m
and for any multi-integer ,
m , the following holds:
u D dx u D dx
0
h
where uh
phuh uhk . K
K K D uh dx u dx
(2.11)
According to Lemma 1.1, in order that u0
H m , the fulfillment of the following condition
is necessary and sufficient
H 0m , , m K K D u dx u D dx 1 h h K K
lim h0
u dx 1
u D dx 0
0
This condition is equivalent to the condition
, m, H0m
m
According to Lemma 1.2, the operator maps H 0
lim , ph u h 0 . H h0
(2.12)
T
onto the space T . Thus the necessary ˆ
result follows from (2.12) and from Lemma 2.1, in which it is assumed:
HT , N H , h phuh .
M
ˆ
Let us consider the approximations of the space H 0m . Let us the sequence phu h weakly converges in space V , i.e. relationship (2.11) holds. According to Lemma 1.1, in order that
u0
H 0m , the following condition is necessary and sufficient , m,
m
Since the operator maps H
H m
onto
lim , ph u h h 0
0.
(2.13)
T 0 (see Lemma 1.2), the necessary result follows ˆ
from (2.13) and from Lemma 2.1 where it is assumed that:
M
H T
HT , N H 0 , h phuh . ˆ
The following theorems determine two sufficient signs of the external FE approximations of Sobolev space.
Theorem 2.2 The FE approximations Vh , ph , r h are external approximations of the space H m -and space
H 0m respectively- when, in some family of spaces Yh H m -and Yh H 0m respectively-, there exists a sequence ph vh Y h such that
ph uh X h
lim p hu h p hv h
h0
H T
0 .
(2.14)
Proof - According to Lemma 1.1, since Yh H m -and Yh H 0m respectively-, for any H 0m
-and H m respectively- it holds the following , m, ph vh Yh
Let us consider the approximations of the space H m
.
, p h v h
H T
0.
(2.15)
Let us phuh be a sequence that
converges weakly in the space V . Then, the sufficiency of the condition (2.14) follows from (2.12), (2.15) and from Lemma 2.2 in which it is assumed that: M
HT , h ph uh , gh ph vh .
When approximations of the space H 0m are considered, the sufficiency of condition (2.14) can be proven in a similar way.
The trace operator
, defined by (1.7), can be written as a direct sum of j-order trace operators
m 1
j
j 0
where
m 1
j ij rK , l , i K K K i r r, l 0
0 j m 1
1, i j 0, i j
ij
(2.16)
The boundary operators in (1.8) can be presented in a similar way
m
j ,
m
j 2 m 1
where m
j
K
Kr Kr ,l i 2 m 1
j 2m 1 1, i j ij 0, i j m
The space from (2.2) can be represented as follows H T
m 1
H
T, j
j 0
where
,
ij rK , l,, i
(2.17)
H T , j
m 1
L
2
ij
K
Kr Kr ,l i 0
K r , l .
1, i j ij 0, i j
Subspaces -see characterizations (2.3) and (2.4)- c an be represented as follows H ˆ
m 1
H ˆ
j
j 0
H 0 ˆ
m 1
H ˆ
0, j
j 0
where H 0, j ˆ
H j ˆ
f H T , j | r , l M r , l
f H 0, j | ˆ
f
0
f r1, l , j
j
1 f r2, l , j (2.18) (2.19)
Theorem 2.3 Let us Vh , ph , r h be a family of FE approximations of the space H m . Let us represent the set I of values of the index j , 0 j m 1 as a sum of its not intersecting and not empty subsets I I1 I 2 .
Approximations
Vh , ph , r h are external approximations of the space H m -and the space
H 0m respectively- if the following holds:
For any j I 1 , there exist families of spaces Yh j H j 1 - and Yh j H 0j 1 respectively- and sequences ph vh j Y hj such that
ph uh X h
lim j phu h j p hvh j
h 0
H T
0,
(2.20)
0,
(2.21)
where j is the trace operator (2.16).
For any j I 2
g
gh j Ghj , ph uh X h
j h
, j ph u h
H T
where Gh j is a dense family of closed subspaces of the spaces H j -and H 0, j respectively-. ˆ
ˆ
Here spaces H j and H 0, j are defined by the characterizations (2.19) and (2.18). ˆ
ˆ
m
Proof – Let us consider approximations of the space H
.
Let us the sequence phuh
converges weakly in the space V , i.e. the relationship (2.11) holds. According to Lemma 1.1, in order that u0
H m , it is necessary and sufficient that the following condition is fulfilled lim , ph u h
, m, H 0m
h0
H T
0.
(2.22)
If the representations (2.16) and (2.17) are taken into account, the inner product in (2.22) can be rewritten as follows
, ph uh
HT
m 1
2m j 1 , j p hu h
j 0
H T
.
(2.23)
The relationship (2.22) will be fulfilled if, for any j I and , m
lim 2 m j 1 , j phuh
h 0
H T
0.
(2.24)
Let us j I 1 . Because of the inclusion Yh j H j 1 , the following relationship holds
,
m,
2 m j 1 , j phu h
H 0m , phvhj Yhj
H T
0
(2.25)
Thus, when j I 1 , then (2.24) follows from (2.20), (2.25) and Lemma 2.2 in which it is assumed that M HT , j , h j ph uh , gh j ph vhj . When j I 2 , the validity of (2.24) follows from (2.21) and Lemma 2.1 in which it is assumed that M HT , j , H H j , h j ph uh . ˆ
When approximations of space H 0m are considered, the result can be proven in a similar way.
The Theorems 2.1-2.3 can be applied not only to justify well-known non-conformed finite elements (in Section 7.1 it is done for the elements of Wilson, Crouzeix -Raviart, Morly, etc.), but also to develop novel highly efficient finite elements and approximations. Some methods of construction of such finite elements are elaborated in the next Section.
3. Types of finite elements for external approximations and their construction Finite elements conformed by a subspace Element definition Let us consider how to construct FE that meet the necessary and sufficient sign established in Theorem 2.1. There are two types of functions in the formulation of the theorem:
Functions g h from the family of enclosed subspaces Gh
H (or Gh H 0 ). ˆ
ˆ
The
domains of determination of the functions are boundaries of the elements. These functions are referred to as boundary functions
Functions phu h from the family X h of FE spaces. The restrictions of the functions phu h onto the FE K belong to the spaces H m ( K ) but the functions itself belong to the space L () only 2
Let us Gh be a family of finite-dimensional spaces. These spaces will be referred to as the spaces of boundary functions. They are elaborated in Section 5. Here, to satisfy the requirements of the Theorem 1.1, the properties of the approximating functions from the spaces PhK
H m K are
established. The types of finite elements that meet the conditions of the theorem are defined, as well as the method of their construction is develop ed.
Let us consider the approximations of the space H m . Using the designations introduced in the Section 1, (2.10) can be rewritten as follows
g h Gh H , ph uh X n ˆ
gh ,
ph uh H
m 1
T
K
K r Kr ,l j 0 K r ,l
m 1
g r ,l j 0 r ,l
g r , l , j r , l , j p d K
1 1 r , l , j r , l , j
K
p
1
K
(3.1)
g r2, l , j r2, l , j p 2 d 0
where K 1
p
, K r M K r , K r , l M K r , l , r , l M r , l P1 H m K 1 ,
p
2
P 2 H m K 2 ,
and where 1
K and K 2 are the adjacent FE
K g r , l , j is a restriction of a function from the space Gh onto the smooth face section K r , l
of the face K r
Because of the inclusion Gh
H and of the characterization (2.4), restrictions of the functions ˆ
from the space Gh onto the boundary
of the domain equal zero, and so are the
corresponding integrals in the expression (3.1).
Let us assume that basic functions of the space Gh H have local supports - they are the subdomains K r , j of the common face
r K r1 K r 2 of the adjacent FE
K and K 2 , with not
0 (it is also possible that
r ). Further it will
empty multitude of internal points K r , j
K r , j
1
be shown that such choice of the basis in the space Gh leads to such basic functions of the finite element space X h of that also have local supports in the domain .
Let us GrK , j designates a finite-dimensional space generated by the basic functions grK ,l ,t whose supports belong to the boundary K of the element K . Because Gh
H , then, according to ˆ
the characterization (2.4), every basic function g r , j , t is a pair g r , j , t
g 1r , j, t , g r2, j, t whose
components satisfy the relation g 1r , j , t
j
1 g r2, j, t , 0 j m 1 .
(3.2)
When the approximations of the space H 0m are under consideration, then the inclusion
Gh
H 0 takes ˆ
place. In this case, due to the characterization (2.5), the boundary basic
functions with supports on the boundary of the domain must be added to the functions whose components satisfy the relation (3.2) (see Fig. 3.1).
From (3.1) it follows that
pi P i H m K i , g ri , j , t Gri , j , i 1, 2;
g1, , 1, r j t
r
1 j p
g r2, j , t r2, j p 2 d 0
r
where rK, j
rK , l , j . K r ,l
0 j m 1 (3.3)
Let us consider FE K 1
and adjacent elements
N K 2 ,..., K (see Fig. 3.1). Let us G K
designates a space of boundary functions of the element GK
0 j m 1,
GrK , j ,
K r M K r .
j , K r
From (3.3) follows:
pi P i H m K i ,
g ri , j , t G i ,
1 i N , 0 j m 1,
g
1 1 r, j,t r , j
1
p
2 k N
(3.4)
g rk, j , t rk, j p k d 0
K 1
For (3.4) to hold there must exist such functions in the spaces Pi ,1 i
N that meet the
following conditions for some arbitrary numbers r , j , t
g ri , j , t ri , j pi d
r , j ,t .
(3.5)
1
K
For i
1 , the integrals in the left hand side of (3.5) are continuous linear forms (functionals) on
1 the space H m K 1 (further, the designation K is used instead of K ). Let us denote these
integrals through i p and let us rewrite (3.5) as follows
i p where M
i ,
1 i M ,
card K , and where K is a set of the functionals (3.5).
The structure of the FE basis is determined b y the following theorem.
(3.6)
Theorem 3.1 Let us the functionals i ,1 i
M from (3.5) be linearly independent on the space P K . Then:
P K can be represented as a direct sum of two subspaces P
K
P P Z ,
(3.7)
where P Z
p P K | i p 0, 1 i M
(3.8)
K and P is a complement to P Z , isomorphic to the factor-space P P Z .
In the space P there exists a basis
p , 1 i M that
satisfies the following
j
condition i pk
1, i k (3.9) 0, i k
Any element p P K can be uniquely expressed as follows
p
M
i 1
where k p , 1 k
i p pi
N M
p p
Z k
k
N M are some coefficients;
p , 1 k
RM , defined by the formula F p
1 p ,..., M p
is surjective. If the opposite is assumed, then there is a vector
such that
(3.10)
k 1
of the space P Z .
Proof - The mapping P K
,
1,..., M R M , 0
Z k
N M is a basis
M
p 0 i
i
i 1
K for any p P , i.e. the functionals i are linearly dependent, which contradicts the initial
assumption.
Canonical basis in the space R M consists of vectors f k , 1 k
M that have a k -component
equal to one and the other components equal to zero. Due to the surjectivity of the mapping F , for any f k there can be found such an element pk P K that
F pk
f k .
Then there exists a set of linearly independent elements p1 ,..., p M
P K that satisfy the
condition (3.9).
Let us P designates a space generated by the elements pi , 1 i M . Then the subspace P Z K defined by (3.8) is a closed complement of the subspace P . Any element p P can be
expressed as follows
p
p1 p2 ,
where M
p1
i p pi P i 1
M
p2
The representation p
p i p pi PZ i 1
p1 p2 is unique because, if 0 p1 p2 , where:
p1
P ,
p2
PZ
then M
p2
ci pi i 1
k p2
M
c p
i
k
i
i 1
From latter follows that
ck i.e. p2
k p2 0, 1 k M
0, p1 0 . Therefore, (3.7) and (3.10) are exact.
The functionals i , 1 i
M associated with FE boundary are referred to as boundary degrees
of freedom; their multitude is denoted by K . The coefficients k , 1 k
N M are
referred to as internal degrees of freedom; their multitude is denoted by AK .
The requirement of linear independence of the set of functionals i p on the space P K is equivalent to the requirement of P K -multi-solvency of the multitude K of boundary degrees of freedom that can be formulated as follows: in the space P K there exists at least one function that satisfies the conditions (3.6) for any set of numbers i . When dim P K definition
of
card K , the
P K -multi-solvency is transformed into the classic definition of
-unique-solvency of the multitude
PK
K (see [10, 27]). When dim P K >card K the multitude
K is P K -poly-solvent and P -unique-solvent. In this case there can exist several subspaces P
PK that provide
P -unique-solvency of the multitude
K .
Now a general definition of a conformed by subspace FE can be formulated. A conformed by subspace FE is a quartet K , G K , P K , P , where:
n
K is a closed domain in R with not empty multitude of internal points and sectionally
smooth regular boundary
P K is a finite-dimensional space of functions determined on the domain K
G K is a finite-dimensional space of functions determined on the boundary K of the domain K
P is a subspace of space P K on which the multitude K of boundary degrees of freedom is PK -unique-solvent
Construction of basic functions of an element
Let us consider how to construct the basic functions of spaces P and P Z . Let us P
H m H
be some finite-dimensional space, that below will be referred to as initial, and let us
pk , 1 k N be its basis, i.e.: p P
N
p
bk pk , k 1
where bk are some coefficients.
(3.11)
The following relations take place N
i p
bk i pk , 1 i M k 1
and they can be rewritten in a matrix form T
Rb
T ,
(3.12)
where:
N
R is a matrix with dimensions M
b is a vector of coefficients b
is a vector of boundary degrees of freedom of FE
N
bk k 1 M
i i 1
Then (3.12) can be rewritten as follows RbT
R Z bZ T T ,
(3.13)
where:
R is a block with dimensions M
M
R Z is a block with dimensions M
N M
Let us the block R be non-singular. Such a block does exist if the rank of the matrix R equals to M , i.e. if the multitude of boundary degrees of freedom of FE is P-multi-solvent. Multiplying 1 (3.13) from the left by the matrix R , the expression for the vector b through the vectors b Z
and is obtained T
b
R1 T R1R Z bZ T .
(3.14)
Now the representation (3.11) can be rewritten as follows p
where
M
1 pk k 1 ;
1bT 2b Z T ,
(3.15)
N
2 pk k N M 1 .
Substituting (3.14) into (3.15), one obtains
p P K
1 R1 T 2 1 R1 R Z bZ T
p
M k p pi rik k 1 i 1 M
N M
j 1
b p j M Z j
M
M
pl
l 1
k 1
Z
rlk r kj
(3.16)
where:
r ik , 1 i, k
r kj , 1 k Z
M , are the elements of the matrix R1
M , 1 j N M are the elements of the matrix R Z
One can see that the element p P
P K has been represented by the elements of the new
basis. When (3.10) and (3.16) are compared, it is apparent that
pk
M
pr
i ik
, 1 k M
i 1
are the basic functions of the space P , while Z j
p
M
M
l 1
k 1
p j M pl rlk rkjZ , 1 j N M
are the basic functions of the space P Z .
M
Thus, to find the basic functions pk
k 1
,
N M
p 1 Z j
j
N
one has to choose a basis pi i 1 of the
initial space P and then to linearly transform it. Herewith, the initial space P and the space PK coincide.
A sufficient condition for existence of an element basis The Theorem 3.1, that determines the structure of FE basis, is based on the assumption of linear independence, or PK -multi-solvency, of the multitude K of boundary degrees of freedom of FE on the space P K
H m K of approximating functions of FE. This assumption implies the
existence of the representation (3.10). The question remained is whether it is always possible to determine the space P K for a given multitude
Let us card K
K that provides P K -multi-solvency of K .
M , dim P K N . Further in the text, dimensions N of the space PK are
varied. For the purpose of differentiation between the spaces P K of different dimensions, the parameter h is introduced into the designation: instead of P K one writes PhK , assuming that h
1 N .
It is apparent that:
When N
M , multitude K cannot be multi-solvent on
When N
M , PK -unique-solvency of K might take place
Multi-solvency of K can only occur when N
M
K
Ph
However, there can be invented the examples when the multitude
K does not become multi-
solvent even if h 0 . In other words, if there are no requirements to the content of the space K Ph , then however big is the dimension N of the matrix R , it does not guarantee the equality of
the rank of the matrix to the number M .
The following theorem defines a sufficient condition for the existence of such h0 beginning with, for which K is PhK -multi-solvent for all h h0 .
Theorem 3.2 For the multitude
K of boundary degrees of freedom of FE to be
PhK -multi-solvent, beginning
with some h h0 , it is sufficient that the family PhK is dense in the space H m K .
Proof - Let us i p , 1 i
M be a multitude of the functionals (3.5) that are the boundary
degrees of freedom of FE. Let us show that these functionals are linearly independent on the space H m K due to the linear independence of the basic functions g r , j , t of the space G K of K
boundary functions of FE.
Trace operators rK, j
K r ,l , j
K r , l
map the space H m K onto the space
,
0 j m 1
H
m j 1 2
K , r l
K r ,l
which is densely enclosed into the space L2 K r . Let us assume that the functionals (3.5) are linearly dependent on H m K , i.e. there exists such a vector that for any u K
r , j , t R M , 0
H m K
m 1
j 0
Kr
t
r , j , t
g rK, j , t rK, j u K d
j0
K r
m 1
Kr K r
r , j ,t
grK, j , t rK, j uK d
t
0 .
From this follows the existence of the vector 0 for which
r , j, t
g rK , j , t 0
t
That contradicts the assumption about linear independence of the basic functions g rK , j , t .
Now let us assume that the family of the spaces PhK is dense in H m K but there is no such h0 beginning with, for which the multitude of the functionals i p , 1 i independent on the spaces
K
Ph . It means that for any
M becomes linearly
h there exists a vector
1,... M RM such that p PhK
M
p 0 . i
i 1
i
(3.17)
From (3.17) and Theorem 1.2 from [33] it follows that the family PhK is not dense in the space
H m K . The achieved contradiction proves the required result.
Virtual finite element In Section 3.1.1, a conformed by subspace FE was defined as a union K , G K , P K , P of four multitudes, where:
K is the element’s domain
G K is a space of boundary functions
P K is a space of approximating functions of FE
P is a subspace of the space P K on which the multitude K of boundary degrees of freedom is P -unique-solvent
When constructing basic functions of the element, according to the procedure described in Section 3.1.2, it is necessary first to introduce some system of basic functions that generates the initial space P K . Then one has to choose a subspace P
P K on which the multitude K of
boundary degrees of freedom of FE is P -unique-solvent. But in the space P K there can exist a finite multitude of subspaces P on which the multitude of boundary degrees of freedom is P -unique-solvent. Any of such subspaces P of this multitude defines its corresponding FE
K, G
K
, P K , P .
If dim P K
N , dim G K
M and N M , then the maximum number of subspaces P
that can be extracted from the space P K is equal to the number of combinations M C N
To extract a subspace P M
N ! M ! N
M !
.
P K one needs to find a non-singular block R with dimensions
M from the matrix R in the system (3.12). It can be performed by the permutation of the
columns in the matrix R (and respectively the elements of vector of the initial system of basic functions).
As it will be shown in Section 4, the convergence rate of the approximations is determined by the approximation qualities of the space P K and does not depend on the choice of the subspace P . The choice of P can only influence the magnitude of the constants in the estimates of the approximation error. Thus, all conformed by subspace FE, characterized by the same multitudes K K , G and P K , possess equivalent properties in terms of the convergence order.
In relation to the above said, it makes sense to introduce the virtual FE as a triplet K , G K , P K , which contains a finite multitude of conformed by subspace FE conformed by subspace FE have the same multitudes
K, G
K
K, G
K
, P K , P . All these
, P K but differ each other in
subspaces P .
Homogeneous and composite finite elements
Let us consider a trivial case when the equations (3.4) are fulfilled
p i P i ,
g ri , j , t G i , 1 i N ,
i
i
i
g r , j , t r , j p d
0
0 j m1 (3.18)
K 1 1 From (3.18) follows that for the element K K , the space P K is a finite-dimensional subspace
of the kernel Z
H m K of the system of M functionals i p
g
K r , j ,t
rK , j pd ,
1 i M .
(3.19)
K
A homogeneous FE is defined as a triplet K , G K , P Z
K
, where:
K is an element’s domain
G K is a finite-dimensional space of boundary functions
P Z K
Z is a finite-dimensional subspace of the kernel of the system of the functionals
(3.19)
Let us consider the procedure for the construction of the basic functions p Z k , 1 k N of a N
homogeneous element. Let us choose a system pk k 1 of basic functions determined on the FE K . The system generates the initial space P which is not necessarily a subspace of the kernel Z
, P Z . Any p P is expressed as N
p
bk pk , k 1
where bk are some coefficients.
(3.20)
By substitution of (3.20) into (3.18), the system of homogeneous algebraic equations is obtained N
b p k
i
k
0, 1 i
M,
k 1
that can be expressed in a matrix form
RbT where R is a matrix with dimensions M
It is assumed that M
N ;
b
0,
(3.21)
N
bk k 1 is a vector of coefficients.
N . If the rank r of the matrix R equals to N , then the system (3.21) has
the trivial zero solution only, and, therefore, the subspace P Z K is empty. If r N then the dimension of the space of solutions of the system (3.21) is N r . In this case a square nonsingular block R1 can be extracted from the matrix R by the permutation of the equations and unknowns. The system (3.21) can be rewritten as follows T
R1b1
R2b2T 0 ,
(3.22)
where R1 is a non-singular block with dimensions r r ; R2 is a block with dimensions
r N
r .
Multiplying the equation (3.22) by the inverse matrix from the left the following complete system of the solutions is obtained
R11R2b2T .
(3.23)
1b1T 2b2T ,
(3.24)
T
b1
Let us rewrite (3.20) as follows p
where
r
N
1 pk k 1 ; 2 pk k r 1 .
Substituting (3.23) into (3.24) one obtains
p P Z K P K
p
2 1 R11 R2 b2T (3.25)
r r bk r pk r pl rlj1 rjk2 k 1 l 1 j 1 N r
From the expression (3.25) it follows that the basic functions p
Z k
N
1 , k
N
N r
of the space P Z K are defined as follows Z k
p
r
r
l 1
j 1
pk r pl rlj1r jk2 ,
where:
rlj , 1 l , j
r jk 2 , 1 j
1
r are the elements of the matrix R11
the elements of the matrix R2 r , 1 k N are
As one can see a homogeneous element has only internal degrees of freedom and does not have any boundary degrees of freedom. It will be shown below that the homogeneous FE generate non-converging external approximations Vh , ph , r h of the space H m on both fixed and variable partitions of the domain . However, these elements are useful when constructing the composite FE described below.
Let us consider two FE:
a homogeneous FE K , G K , P Z K ,
a conformed by subspace FE K , G K , P K , P .
ˆ
ˆ
ˆ
Let us the following relations hold:
K ˆ
K ,
G
G K ,
(3.27)
P Z K
PZ .
(3.28)
P PZ PZ K
(3.29)
ˆ
K
ˆ
(3.26)
K
Then the element K , G K , PC , P , where:
PCK
ˆ
is referred to as a composite FE. A composite FE is a superposition of a homogeneous and a conformed by subspace FE. The equality (3.26) means the same geometry of the elements. Equality (3.27) implies the property K
P Z ˆ
P 0 .
(3.30)
The properties (3.28) and (3.30) provide the expansion of the space P K due to the addition of new functions from the space P Z K . It can be useful for improvement of the approximation ˆ
qualities of FE.
Spaces of finite elements
Let us characterize the spaces of the FE defined above. As it was assumed in Section 3.1, the support of every basic function g r , j , t , from the space Gh of the boundary functions, is either a sub-domain of an inter-element boundary
r , i.e. supp gr , j , t r , or a sub-domain of the
face K r coinciding with the boundary of the domain , i.e. supp g r , j , t
and K r (here
below are used the symbols accepted in Section 3.1).
Every basic function g r , j , t with a support in
r is a pair
gr , j,t
g 1r , j, t , g r2, j, t whose components
1 satisfy the relation (3.2). Herewith, in accordance with Section 3.1, for the adjacent FE K and
K 2 the following holds
p1 P1 ,
g
1 1 r , j, t r , j
p 2 P 2 , 1
p
0 j m1
g r2, j, t r2, j p 2 d 0
r
that can be rewritten as follows r1, j , t p1
r2, j, t p 2 ,
where: 1r , j , t p1
r
r2, j , t
2 2 2 g r , j , t r, j p d
g
p 2
r
1
r, j, t
1r, j p1d
(3.31)
1 are the boundary degrees of freedom of the elements K and K 2 respectively. Then the spaces
m
X h of conformed by subspace FE that approximate the space H
characterization
are
defined by the
X h
v p K K P K | K
g1r , j , t r1, j p1d
r
g r , j , t Gh H ˆ
g r2, j , t r2, j p 2 d , r
0
j m 1
The characterization of the space X 0 h that approximates the space H 0m is
X 0 h
v p K K P K | g r , j , t Gh H 0 K g1r , j , t r1, j p1d g r2, j, t r2, j p 2d , ˆ
r
r
g r , j, t
K r, j
K
p d
0,
K
r
0 j m 1
Basis of the space Gh generates the multitude h of boundary degrees of freedom of the space of FE. Every degree of freedom from the multitude h is a pair r1, j , t , r2, j , t whose elements are determined by the relations (3.31). Any FE also possesses a possibly empty multitude of internal degrees of freedom: B K
i , 1 i N K M
where (see Section 3.1.1): N K M
dim P K card K dim P
The multitude of internal degrees of freedom of FE space is denoted through Bh : Bh
B K K
K
, card Bh N h M h
where N h
N
K
, M h dim Gh .
K
Any element from the space X h of conformed by subspace FE can be uniquely defined by multitudes h and Bh of boundary and internal degrees of freedom.
The basis of the space X h consists of the basic functions of two types:
wr , j , t ˆ
ˆ
ˆ
X h
ri , j , t wr, j , t ˆ
ˆ
ˆ
g
i r, j, t
r
ˆ
ˆ
ˆ
1i 1 when 0 otherwise
Z
wk
ri , j wr, j , t |K i d r
(3.32)
r , j j, t t , i 1, 2 ˆ
ˆ
X h , ri , j , t wkZ 0,
i
ˆ
1, 2 .
(3.33)
Here the symbols:
r , r of the adjacent FE
r , r identify the inter-element boundaries
j , j identify the orders of the derivatives that participate in the definition of boundary
ˆ
ˆ
ˆ
degrees of freedom
K
t , t identify the basic function from the spaces Gr , j ˆ
The equalities r r , j j , t t mean the coincidence of: ˆ
ˆ
faces
r r , ˆ
ˆ
orders of differentiation
IDs of basic boundary functions
Basic functions of the space X h are constructed from the basic functions of FE in the following way. Let us:
1 , 2 h is a boundary degree of freedom
g is a basic boundary function that generates
K and K are adjacent FE on whose common face is defined the function g
pi
1
2
Pi , i 1, 2 are the basic functions from the spaces Pi of the elements
i
K that
correspond to the boundary degree of freedom
Then the function w X h , defined by the condition
pi x , w 0,
x Ki , x
i
1, 2
K i ,
is the basic function of the space X h corresponding to the boundary degree of freedom and satisfying the conditions (3.32).
Let us p Z
P Z designates a basic function from the space P Z of the element
Z to the internal degree of freedom . Then the function w
K corresponding
P Z defined as:
p Z x , x K w x K 0, Z
is the basic function (3.33) associated with the internal de gree of freedom .
(3.34)
Thus, the support of any basic function wr , j ,t
X h consists of two adjacent FE (see Fig. 3.2). A
Z support of a function w Z k is a single FE. Functions wk and wr , j ,t are linearly independent and
they generate the whole space X h .
Let us now examine the spaces of homogeneous FE. For such FE the basic functions of the space
Gh of boundary functions do not generate boundary degrees of freedom of FE. The space X h of
homogeneous FE is characterized as follows
X h
v p K K PZK | K
g 1r , j , t r1, j p1d
r
g r , j , t Gh H
g r2, j , t r2, j p 2d 0,
ˆ
0 j m 1
r
Any element from the space X h can be uniquely expressed through the multitude
Bh
BK bl , 1 l Lh K
of internal degrees of freedom of elements. Herewith dim X h Lh .
If p Z
P Z K designates a basic function of a homogeneous FE corresponding to the internal
degree of freedom b , then the basic function w Z X h is defined by conditions (3.34). Its support is a single FE.
A composite FE is a superposition of a conformed by subspace FE and a homogeneous FE. Therefore, the space X hC of the composite FE is a direct sum of the corresponding spaces X hC
X h X h .
(3.35)
Herewith dim X hC dim X h dim X h .
A function from the space X hC is uniquely defined through the multitudes Bh and Bh . Basic functions from the space X hC are constructed from the basic functions of the conformed by subspace and homogeneous FE.
Figure legends for Section 3:
Fig. 3.1 To the definition of boundary functions with local supports
Fig. 3.2 Supports of basic functions from FE spaces
4.Error estimates and convergence of external approximations of Sobolev spaces Operators of approximations Let us consider a conformed by subspace FE K , G K , P K , P , where:
dim P K
N , dim G K
and let us introduce the function v H m K ,
v:K
M
R.
Let us
AK v
K v SK v
(4.1)
designates a P K -approximant of the function v on the domain K . Here, the mapping
K : v p P is defined as follows M
K v i v pi ,
(4.2)
i 1
and the mapping Sk : v SK v
p PZ N M
v p i
Z i
i 1
is the best approximation in the space L2 K of the function v by the shift of the subspace P Z
P K , i.e.
p P Z
v K v S K v, p 0, K 0 .
(4.3)
Because the subspace P Z has the basis
p , 1 i N M , Z i
then the equation (4.3) is
equivalent to the system
v
K
v
SK v, p Z i 0, K 0
from which the coefficients i v , 1 i
1 i N M
(4.4)
N M of the best approximation are determined.
The approximation (4.1) of the function v is unique by virtue of the P -unique-solvency of the multitude K of boundary degrees of freedom of FE on the subspace P and because of the linear independence of the system of equations (4.4). Because the boundary degrees of freedom of FE i , 1
i M are determined on the space H m K , then the domain of determination of
the operator of P K -approximation is the space H m K , i.e. dom AK P
K
H m K . On the space
dom AK the operator of P K -approximation is identical operator, i.e. p P K
AK p
p
(4.5)
Let us space X h is the space of the conformed by subspace FE with the multitude of boundary degrees of freedom h and with the internal degrees of freedom Bh . Herewith
dim X h Nh , card h M h , card Bh Lh .
With an arbitrary function v H
m
,
v:
R
Ah v
Here
is associated its
hv Shv .
X h
-approximant (4.6)
M h
h v j v wj ,
(4.7)
j 1
where w j , 1 j
M h are
the basic functions (3.32), and i
h , 1 j M h are the
boundary degrees of freedom of the space X h .
The operator S h is defined by the relation
v v S v, w 0, h
h
Z i
0,
1 i Lh ,
(4.8)
where w Z i , 1 i Lh are the basic functions (3.33).
The function
h v X h is characterized by the following property j
h v j v , 1 j
Mh .
Because the basic functions w Z i , 1 i M of the space X h coincide with the basic functions p Z i ,1 i
N M of the spaces P Z P K , and the support of the function p Z i is a single
element, then the system (4.8) is equivalent to the finite set of systems of sort (4.4) formulated for all K . Therefore, the function S h v is characterized by the following property
i
Bh , Sh v X h
and, for any function Ahv
i Sh v
i v , 1 i Lh ,
X h , the following properties take place: j Ah v
j v , 1 j M h
i Ah v
i v ,
1 i Lh
The domain of determination of the operator Ah of X h -approximation is the space H m , i.e.
dom Ah H m .
Between the operators AK and Ah there exists the dependence determined by the following theorem.
Theorem 4.1 Let us v be an arbitrary function from the space H m . Then restrictions v K belong to the space H m K and
K
Ah v K
AK v
K
.
Proof - Relations (4.9) follow from the definitions of the multitudes
Let us define the approximation operators for a homogeneous FE *
(4.9)
h and Bh .
K, G
K
, P Z K . The operator
K
AK of P Z -approximation can be defined as an operator of orthogonal projection in the space 2
L
K of the function v H m K onto the space p P ZK
If
p , Z k
P Z K
H m K , i.e.
v A* v, p 0, K
K
0.
(4.10)
1 k N * is a basis of the space P Z K , then the relation (4.10) is equivalent to the
following equation system
v A* v, p 0, Z k
K
K
0,
1 k N* ,
(4.11)
where *
N
*
AK v
bk v p Z k . k 1
The uniqueness of P Z K -approximation of the function v follows from the linear independence of the system (4.11).
The following relations also hold
dom AK * L2 ( K ) H m K ,
p P ZK
*
AK p
p
Let us X h* be the space of homogeneous FE. X h* -approximant of function v is associated with an arbitrary function v H
m
,
v:
R . The operator of X h* -approximation
Ah* : v
Ah*v X h*
is defined by the relation
v A*v, w* 0, h
i
0,
1 i Nh* ,
where wi* , 1 i N h* are the basic functions of the space X h* . Because the support of each basic function wi* is a single FE, then any function Ah*v X h* is characterized by the following property bi Ah v *
bi v ,
1 i N h* .
Let us consider a composite FE
K, G
K
, PC K , P , where PCK P PZ PZ K . The PC K ˆ
-approximant of the function v H m K is AKC v
where the mapping
K v S K C v ,
(4.12)
K : v p P is defined by (4.2), and the mapping C SK : v
p PZ PZ K ˆ
defines the best approximation in the space L2 K of the function v by the shifts of the space P Z
PZ K , i.e. ˆ
v
p P Z PZK ˆ
Let us p Z i , 1
K
v
SK C v, p 0, K 0 .
(4.13)
i N M be a basis of the space P Z , and let us p Z j , 1 j N * ˆ
ˆ
be a basis of the space P Z K . Then the equation (4.13) is equivalent to the system of the equations: ˆ
v v
S KC v, piZ 0, K 0, 1 i N M . C Z * , 0, 1 v S v p j N K K j K 0, K
v
(4.14)
ˆ
ˆ
Herewith the PC K -approximant of the function v is unique because of the P -unique-solvency of the multitude K of boundary degrees of freedom of FE and because of the linear independence of the system (4.14).
The following relations hold:
dom AK C H m K
p PCK
C
AK p
p
(4.15)
Let us X hC be the space of composite FE. For an arbitrary function v H m , its X hC -approximant is:
hv S hC v ,
AhC v
(4.16)
where the operator S hC is determined by the relations
v v S v v S h
C h
v, wiZ
h
C h
v, wiZ
0,
ˆ
0, 1 i Lh 0, 1 i Nh* ˆ
0,
and the operator h is defined according to the representation (4.7). Herewith
dim X hC Nh N h* . ˆ
It is easy to verify that the operators AhC and Ah* possess the property determined by the Theorem 4.1 for the operator Ah .
Let us
Vh , ph , rh , Vh* , ph , rh* , VhC , ph , rhC
be the external FE approximations of Sobolev
spaces, constructed respectively on conformed by subspace FE, on homogeneous FE, and on composite FE. Restriction operators of the approximations are defined as follows: m rh : v H v h
1 ,..., M ,1 ,..., L R N Vh
m * rh : v H vh
1 ,..., N R N Vh*
C m rh : v H vh
h
h
h
*
h
*
h
1 ,..., M ,1 ,..., L , b1 ,..., b N R N h
h
ˆ
*
h
h
N h* ˆ
C Vh
(4.17)
Then the corresponding operators of approximations can be defined as products of extension operator (1.4) and restriction operator (4.17):
Ah
ph rh , Ah* ph rh* , AhC p h r hC .
Convergence of approximations on a fixed partition Convergence properties of external FE approximations of Sobolev spaces on a fixed partition of the domain are defined by the following theorems.
Theorem 4.2 Let us Vh , ph , r h be a family of external FE approximations of the space H m associated with a fixed partition of the domain and built on the conformed by subspace FE. Approximations Vh , ph , r h converge in H m if, and only if, for each FE K the family m of spaces PhK of approximating functions of FE is dense in the space H K .
Necessity - Let us assume that the external approximations Vh , ph , r h converge, i.e.
u H m where ph rh spaces H
m
lim u ph rh u h0
V
0
Ah is the operator of X h -approximation. Then the density of the family
K,
K
Ph in the
K follows from the definition (1.2) of the norm on the space V and from
the dependence (4.9) between operators of X h - and PhK -approximation.
m Sufficiency - Let us determine the X h -approximant of the function u H
Ahu
hu Shu ph rhu
where the operators h and S h are defined by the relations (4.7) and (4.8) respectively. For any K and for any h , the PhK -approximant of a function v K
AK , h v
K
Ahu k
M h
v p K
i, h
i 1
i, h
u K is defined according to (4.1)
M h
Nh
i 1
i, h v K p iZ, h ,
(4.18)
where:
i , h v K , i , h v K are the boundary and internal degrees of freedom of FE that are linear and continuous functionals on H
m
K
pi , h , piZ , h are the basic functions of FE
N h
dim( PhK ) and M h is the number of boundary degrees of freedom of FE
From the family of spaces PhK let us chose such sequences of functions p Kh PhK that ˆ
K
lim v K phK ˆ
h 0
m, K
0.
(4.19)
These sequences exist because, according to the assumptions, the families of the spaces PhK are dense in the spaces H
m
vK
K . From the triangle inequality one has AK , h vK
v K phK ˆ
m, K
phK Ak , h vK
m, K
which yields
u here
Ahu
V
u uh ˆ
ˆ
V
uh Ahu ˆ
V
,
m ,K
uh ˆ
p ˆ
K h
K .
,
K
The Theorem will be proven if it can be established that
lim phK AK , h vK ˆ
h0
m, K
0.
(4.20)
According to the Theorem 3.1, the element p Kh PhK can be uniquely expressed as follows ˆ
K h
p ˆ
M h
p p i 1
i, h
ˆ
K h
i, h
Nh M h
i 1
i, h p Kh piZ, h .
ˆ
(4.21)
From (4.19) follows:
lim i , h v K phK 0, ˆ
h0
lim i , h v K phK 0 ˆ
h0
Therefore i , h phK ˆ
i , h phK ˆ
, v , v K
i h
K
i h
and the relation (4.20) is the implication of the representations (4.18) and (4.21).
m Let us note that the density of the family PhK in the space H K is the sufficient condition for
the PhK -multi-solvency of the multitude of the boundary degrees of freedom of FE (see Theorem 3.2), as well as it is also necessary and sufficient condition for the convergence of the approximations on a fixed partition of the domain (see Theorem 4.2). Therefore, FE approximations of Sobolev spaces, constructed on conformed by subspace FE and a fixed
partition of the domain, can be external but non-converging. The following Theorem shows that external approximations constructed on homogeneous FE are always non-converging.
Theorem 4.3
External FE approximations Vh* , ph , r h* of the space H m associated with a fixed partition
of the domain and built on homogeneous FE are non-converging in the space H m .
Proof - Let us consider an element K . The family of subspaces PhK is dense in the space
H m K if, and only if, any continuous on H m K functional that equals zero on PhK , equals zero on H m K also (see Theorem 1.2 from [33]). For homogeneous FE functionals (3.18) equal zero on the elements from PhK . Obviously, functionals (3.18) do not turn to zero for any function v H m K and, therefore, families of the spaces PhK are not dense in the spaces
H m K .
Theorem 4.4
Let us VhC , ph , r hC be a family of external FE approximations of the space H m associated with a fixed partition of the domain and built on composite FE. For the approximations
V
C h
, ph , r hC to be converging in the space H m it is necessary and sufficient that families
K Ph of the spaces of approximating functions of conformed by subspace FE contained in the
spaces PCK , h of composite FE are dense in the spaces H m K .
Proof – Proof of this theorem follows from (3.29), (3.35), Theorem 4.2 and from the definition
of the operator of X hC -approximation (see formula 4.16).
Convergence of approximations on a family of partitions Some definitions Let us h be a family of partitions of the domain , which satisfies the conditions of regularity (2.1). In classic FEM, the partition h of the domain is usually conducted in the way that all FE K h are linearly equivalent to some initial FE K , also referred to as common FE (see, for ˆ
example, [10]). In this case any element K h can be obtained from the common element K ˆ
through the linear mapping
F : x Rn ˆ
F x Bx b R n , ˆ
(4.22)
ˆ
n where B is an invertible matrix of dimensions n n ; b is a vector in R . The concept of
linearly equivalent FE is introduced to obtain the approximation error estimates.
Let us P K be a space of functions p : K space of functions g : K
R defined on the domain
K
K , and let us G be a
R defined on the boundary K of the domain
K . Because the
mapping (4.22) defines one-to-one correspondence between the points of the domains K and K ˆ
then the following spaces can be associated with the spaces P K and GK P*K
p : K R p p F 1 , p P ,
G*K
g : K R g g F 1 x , x K , g G
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
where P and G are the spaces of functions defined on the domain K and on the boundary K ˆ
ˆ
ˆ
ˆ
respectively. Then, due to the linearity of the mapping F K
P*
P K , G*K G K .
So the relations between the points x K and x K and between the functions p P and ˆ
ˆ
ˆ
p P K , g ˆ
ˆ
G and g G K are the following ˆ
p P p p F 1 P K g G g g F 1 x G K , x K
x
ˆ
K x F x K ˆ
ˆ
ˆ
ˆ
ˆ
(4.23)
ˆ
ˆ
ˆ
From the relations (4.23) it follows that p x
p x
x K , p P,
g x
g x
x K , g G
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
Let us K , G K , PK , P be some arbitrary conformed by subspace FE that is characterized by the ˆ
ˆ
ˆ
ˆ
condition
diam K 1 . ˆ
(4.24)
For any element
K, G
K
K, G
h , through , P K , P , K
ˆ
ˆ
K
, P K , P let us designate some ˆ
ˆ
initial element for which:
K K 1 P p : K R p pF , pP K K 1 G g : K R g g F x , x K , g G K
F K ˆ
ˆ
ˆ
ˆ
(4.25)
ˆ
ˆ
ˆ
Moreover, let us the mapping (4.22) be a similarity mapping, when B is a diagonal matrix with the following diagonal elements:
hK , 1 i n, hk diam K .
bii
The similarity mapping maintains the normal to FE bound ary.
The following relations take place for the similarity mapping:
jK v hK j K j v ˆ
ˆ
d hK n 1 d 0 j m 1
(4.26)
ˆ
K where j is the operator of differentiation of j -order along the normal to FE boundary; d is
an elementary measure of a n 1 -dimensional surface in R . n
Let us modify the definition of the boundary degrees of freedom of FE by adding the multipliers n j 1
hK
: v
hK n j 1
K r
K
K
g j vd ,
where g K is a basic function of the space Gh of boundary functions. Let us note that the multipliers do not change the definition of FE given in the Section 3.1.1. The modification allows to prove the important relation between the operators of P K - and P K -approximations, ˆ
which later will be used in establishing the convergence of the approximations.
Theorem 4.5
K, G
Let us K , G K , P K , P , ˆ
ˆ
ˆ
ˆ
K
, PK , P be two similar conformed by subspace FE. If
pi , 1 i ˆ
M,
p Z j , 1 j
N M
p Z j , 1 j
N M
ˆ
are the basic functions of the element K and if: ˆ
pi , 1 i
M,
are the basic functions of the element K , then for the operators of approximation AK and AK ˆ
the following relation takes place ^
AK v
AK v ˆ
ˆ
for all functions v dom AK , v dom AK that are in correspondence ˆ
ˆ
v
ˆ
dom AK v v F 1 dom AK . ˆ
ˆ
Proof – The P K -approximant of the function v is (see Section 4.1) M
AK v
i v pi i 1
where coefficients j v ,1 j equations as in (4.4):
N M are
N M
v p j
Z j ,
j 1
determined from the system of linear algebraic
N M
j
j 1
M v i v pi , pt Z , 1 t N M . (4.27) i 1 0, K
v p , p 0, K Z j
Z t
Taking into account that F is the similarity mapping which maintains any normal orthogonal to the FE boundary, and using relations (4.26), one obtains:
i v
hK n j 1
g K Kj vd
hk n j 1
K r
g K hK j Kj vhKn 1 d ˆ
ˆ
ˆ
ˆ
ˆ
Kr
g K Kj vd ˆ
ˆ
ˆ
ˆ
ˆ
i v . (4.28) ˆ
ˆ
K r
ˆ
ˆ
For the initial FE the system (4.27) is: N M
j 1
j v p Zj , ptZ ˆ
ˆ
ˆ
ˆ
0,
M v i v pi , pt Z , 1 t N M . (4.29) i 1 0, K ˆ
K ˆ
ˆ
ˆ
ˆ
ˆ
ˆ
Taking into account the equality (4.28) and the relation
f , g 0, K
fgdx fgh dx h f , g ˆ ˆ
K
n K
n K
ˆ
ˆ
ˆ
K
0, K ˆ
ˆ
the system (4.29) can be rewritten as follows hK n
N M
j 1
j v p Zj , ptZ ˆ
ˆ
0,
K
M hK n v i v pi , ptZ , 1 t N M . (4.30) i 1 0, K
Comparing the systems (4.30) and (4.27) it is apparent that
j v
j v , 1 j N M . ˆ
ˆ
Therefore AK v
M
i v pi ˆ
ˆ
N M
i 1
v p ˆ
j
Z j
ˆ
.
j 1
From this, using the relations (4.23) one can finally obtain M
AK v i v pi ^
ˆ
i 1
ˆ
ˆ
N M
v p ˆ
j 1
j
ˆ
ˆ
Z j
AK v . ˆ
ˆ
Error estimates for piecewise polynomial approximations In the estimates below C designates a constant independent of the parameter h of the family of approximations and considered herewith functions, Pk designates a space of the polynomials of k -order determined on the corresponding domain.
Theorem 4.6
Let us
K, G ˆ
ˆ
K
, PK , P be a conformed by subspace FE. Let us m 1 be the highest order of ˆ
ˆ
normal derivatives that participate in the definition of the boundary degrees of freedom of FE. If, for some whole number k 0 , take place the inclusions:
H K ,
H k 1 K
m
ˆ
P
Pk K ˆ
ˆ
K
(4.31)
ˆ
H m K ,
(4.32)
ˆ
then there exists such a constant C that for all similar FE
K, G
K
, P K , P and for any function
u H k 1 K
u
AK u l , K
hK k 1 C l u k 1, K , K
0 l m,
where AK u is a PK -approximant of the function u and where:
hK here S is a sphere from K .
diam K ,
K
sup diam S
(4.33)
Proof – From the inclusion (4.32), accounting the fact that the operator of P K -approximation ˆ
is identical on the space P K (see Section 4.5), it follows that ˆ
p Pk K
AK p ˆ
ˆ
ˆ
ˆ
p. ˆ
Let us u be a function from the space H k 1 K . Because of the inclusion (4.31), this function ˆ
ˆ
also belongs to the space dom AK ˆ
H m K . The linear mapping ˆ
H K
AK : H k 1 K ˆ
ˆ
l
ˆ
is continuous because of the inclusions (4.31) and (4.32). Herewith, the operators of P K - and ˆ
PK -approximation are in the relation (4.9)
u dom AK
^ AK u
AK u . ˆ
ˆ
Then the necessary result follows from the Theorem 3.1.4 from [10].
Theorem 4.7
Let us there be given a regular family of conformed by subspace FE K , G K , P K , P . Let us for any FE from this family, inclusions (4.31) and (4.32) take place, in which K is substituted by K ˆ
. Then there exists such a constant C that for all FE from the family and for any function 1
u H k
K the following inequality holds u
AK u
l, K
ChK k 1 l u k 1, K ,
0 l m.
(4.34)
Proof - For any element K from the family it can be put in correspondence through the
similarity mapping some initial element K that meets the relations (4.24) and (4.25). Then the ˆ
estimates (4.33), in which constants C vary for different elements, are true. Because the constants C are bounded then the estimate (4.34) follows from the estimates (4.33) and the regularity conditions (2.1).
Theorem 4.8
Let us there be given a family of external FE approximations Vh , ph , r h of the space H m associated with a regular family of partitions h of the domain and built on the conformed by subspace FE. Let us there exists such whole number k 0 when m k allowing the following inclusions to take place
K h
Pk K
P K H m K .
Then, there exists such a constant C independent from h , that for any function u H k 1
the following inequalities hold: 1
2 2 k 1 l u k 1, , u Ah u l , K Ch K u Ah u V Chk 1 m u k 1, where Ahu
0l m
X h is X h -approximant of the function u defined by the relation (4.6).
Proof - Applying the Theorem 4.7 the following estimate is obtained
u
AK u
l, K
ChKk 1 l u k 1, K
Using the relations (4.9) and the inequalities hK
u AK u K
2 l, K
1
2
Ch
k 1 l
K h , 0 l m .
h, K h one obtains the following 2 u k 1, K K
1
2
Ch k 1 l u k 1,
Let us also note that
u AK u K
m, K 2
1
2
u Ahu V .
Convergence of approximations Let us determine the conditions of convergence of external FE approximations associated with a family h of partitions of the domain , i.e. the conditions when the following holds
u H m where ph rh
lim u ph rh u h0
Ah is an operator of X h -approximation.
Theorem 4.9
V
0,
Let us Vh , ph , r h be a family of external FE approximations of the space H m associated with a regular family h of partitions of the domain and built on conformed by subspace FE. If the inclusions
K h
Pm K
PK H m K
take place, then the approximations Vh , ph , r h converge in the space H m .
Proof - Using the Theorem 4.8 the following estimate is obtained
v H m 1
v
ph rh v V Ch v m 1, .
Therefore
lim v ph rhv h 0
V
0.
(4.35)
From the triangle inequality for any h and any v H m 1
u By introducing the function u H
m
function v
phrhv
V
u v
V
v phrhv
V
.
(4.36)
and the arbitrary number 0 , one can determine the
H m 1 that satisfies the inequality u
v
V
2
This is possible due to the density of the inclusion H m
1
H m . According to (4.35),
there exists such h0 that
h h0
v
ph rh v
V
2
.
Therefore, taking into account the inequalities (4.36), it is obtained
u H m
lim inf u ph vh h 0 v !h V h
V
0 .
Theorem 4.10
A family of external FE approximations Vh* , ph , r h* of the space H m associated with a regular family of partitions h of the domain and built on the homogeneous FE does not converge in the space H m .
* Proof - In case of homogeneous FE for any member of the sequence phu h X h the following
functionals become zero (see Section 3.2)
g
K
jK ph uh
K
K
d
0,
0 j m 1,
(4.37)
where g K is a restriction on K of the basic function from the space Gh of boundary functions. However, for any function u H
m
V the
functionals (4.37) do not become zero.
Therefore, according to the Theorem 1.2 from [33], it is impossible to segregate the sequence phu h from the family of spaces X h* converging to the given function u H
m
of the space V .
Theorem 4.11
in the norm
Let us VhC , ph , r hC be a family of external FE approximations of the space H m associated with a regular family h of partitions of the domain and built on composite FE. If the following inclusions take place
K h
Pm K
PC K H m K
then the approximations VhC , ph , r hC converge in the space H m .
Proof – Proof of the Theorem is similar to the proof of Theorem 4.9 because the properties
(4.15) and (4.9) of the operators AhC and AK C of the approximation, determined by the relations (4.12) and (4.16), are identical to the properties (4.5) and (4.9) of the operators Ah and AK for the approximations built on conformed by subspace FE.
5. Approximations of boundary spaces Definitions To construct the finite elements described in Section 3 it is necessary to know how to build families of spaces Gh of boundary functions that are dense in the boundary spaces H , H 0 (see ˆ
ˆ
characterizations (2.5) and (2.6)). The approximations of the spaces H , H 0 by the spaces Gh ˆ
ˆ
are referred to as boundary approximations, and they are examined in this Section.
Here and below C designates various constants independent of parameter h of the family of the approximations and herewith considered functions; Pk designates the spaces of polynomials of k -order, determined on the corresponding domains.
n Let us R be an open domain with sectionally smooth regular boundary , and let us
be a connected patch of the boundary . Let us assume that
can be represented as a
combination of the finite number of domains of th e same dimensions, i.e. N
j , N j 1
and let us assume that each of the domains j can be continuously and one-to-one mapped onto the domain D j of the Euclidean space R x
j t
n 1
:
x ,
t
t1 , ..., t n 1 D j R n 1
where j t C 0 D j ,
1 j N.
(5.1)
It is assumed that v H S if v j H S D j . Herewith:
The map (5.1) is called the local map of the surface
j ,1 j N is called the domain of the local map
D j ,1 j
N is called the domain of parameters of the local map
If one map is sufficient to define the surface , then such surface is called elementary surface. Local maps (5.1) can be built by straightening part j of the surface through the introduction of new coordinates. Herewith the coordinates ti
ti x are
defined in the vicinity of point
x0 in such a way that within these coordinates, part of the boundary j has the equation t n
0 and part of the domain , adjacent to it, is located in subspace t n 0 . The coordinates
ti ,1
i n can be introduced in the following way. Smooth section j is associated with the
Cartesian coordinate system y1 ,..., yn so that j can be defined through the equation
yn
f y,
y
y1 , ...yn 1
(5.2)
and variables x are substituted by variables y . Then new variables t are introduced by the formulas
yi , 1 i n 1, t n yn f y . ti
Further, for simplicity, only elementary surfaces are considered. Norm and semi-norm on the S
space H
are defined as follows
v
s,
v s,
1 2 S 2 2 Dt v 1 grad f dt D s 1 2 2 2 Dt v 1 grad f dt D s
(5.3)
where:
t1 ,..., t n 1 of the map of the surface
D is a domain of parameters t
Dt is an operator of differentiation by variables t1 ,..., t n 1
function f t is defined by relation (5.2)
dt
grad f f dt1,..., f dt n 1
1,..., n 1 is a multi-index
dt1 ,..., dtn 1 ;
The expression d
2
1 grad f dt
defines the n 1 -dimensional measure of the elementary surface.
When s 0 , the expression (5.3) defines the norm on space the L2 induced by the inner product
u, v 0, uvd .
Let us be given a domain
Rn with
sectionally smooth regular boundary
. Let us a
partition of the domain into FE K is defined. In addition to the boundary spaces H T and H T (see characterizations (2.2), (2.4) and (2.5)) the following spaces are defined W
m 1
H K ,
s j 0,
r ,l
K
s j 1
K r Kr ,l j 0
W0 f W | r ,l M r ,l W
where K
f W0 |
1
f r ,l , j
f 0
j 1 f r2, l , j , 0 j m 1
(5.4)
, K r M K r , K r , l M K r , l .
In the characterization (5.4), figures 1 and 2 indicate that the functions are defined on the inter1 2 element boundaries r , l of adjacent FE K and K . Herewith, for every
1 2 are such elements K and K that
r , l M r , l there
r ,l Kr1,l K r2,l .
The norm on the space W is:
g
W
m 1 g rK , l , j K K K j 0 r
1
2 s j 1, K r ,l
r ,l
2 .
The following semi-norms on W are also useful:
g j grK , l , j K K K 0 j m 1 r
1
2 s j 1,K r ,l
r ,l
Spaces W and W 0 are the closed subspaces of the space W .
2
(5.5)
Operators of boundary approximation Let us there be given some partition of the domain into FE K . In Section 3.1 it was assumed that the support of any basic function g r , j , t , 0 j m 1 of the finite-dimensional space Gh of boundary functions is:
either a connected sub-domain K r , j of the common boundary
r K r1 K r 2 of
1 adjacent FE K and K 2 , i.e.
Kr, j
supp g r , j , t r ,
K r , j
0
or a connected sub-domain of the face K r * of FE that coincides with the boundary Kr , j
supp g r , j , t K r* (it is also possible that
Any of the basic functions g r , j , t with support in gr, j,t
where g 1r , j , t
K r , j
r or Kr , j K r * ).
r is a pair
g 1r , j , t , g r2, j, t ,
g r , j, t |K , g r2, j, t g r, j, t |K . 1
2
r
Because of the inclusions Gh
r
H H 0 the components of the pair satisfy the relation ˆ
1
gr , j ,t
ˆ
j
1 g r2, j, t , 0 j m 1 .
Let us introduce the following designations:
(5.6)
, i.e.
g rK , j ,t is the restriction of the function g r , j , t onto the face of the element K
Gr , j is the space generated by the functions g rK , j ,t
K G K is a union of the spaces Gr , j
K
The space Gh of the boundary functions can be represented as follows: Gh
G K ,
K
(5.7)
K
where m 1
G
GrK , j .
K
(5.8)
K r j 0
The assumption (5.6) allows the construction of the spaces of FE whose basic functions have a local support in the domain . With a specific choice of the functions g r , j , t , it also allows to prove the density of the family Gh of the spaces of boundary functions in the space H 0 or H . ˆ
ˆ
It will be shown in the Section 6 that, in order to obtain optimal error estimates of the approximations of variational equations, it is necessary to strengthen the assumption (5.6).
Let us designate through g r , l, j , t a basic function of the space Gh whose support is:
either a connected sub-domain K r , l , j of the smooth section common face
r of adjacent FE K r ,l , j
1
2
K and K :
supp g r , l , j, t r, l ,
K r , l , j
0
r ,l Kr1,l K r2,l of
the
or a connected sub-domain of the smooth section with the boundary
Kr , l , j
K r*, l of
the face K r * that coincides
, i.e. Kr ,l , j K r*,l (it is also possible that
K r , l , j
r , l or
K r* ).
Then the space G K in the expression (5.8) can be represented as follows m 1
G
K
GrK ,l , j ,
(5.9)
Kr Kr ,l j 0
where GrK ,l , j is the space generated by the basic functions grK , l, j ,t .
K
K
Let us designate through Pr , j (and through Pr ,l , j correspondingly) the finite-dimensional spaces of the basic functions g rK , j ,t (and grK , l, j ,t correspondingly) that have a common support K r , j (and K
K
K r , l , j correspondingly). By definition, the spaces Pr , j and Pr ,l , j are not empty. The domains K r , j
and K r , l , j , along with their spaces PrK , j and PrK ,l , j , are referred to as surface elements. Therefore, a surface element can be characterized by the pair of multitudes K r , j , PrK , j or K r , l , j , PrK , l , j .
Let us conduct the partitions r , j , 0 j m 1 of the face K r into the surface elements Kr , j
r , j , 0 j m 1 so that the following conditions hold:
Kr Kr , j ,
for any K r , j r , j the domain K r , j is closed and the multitude of its internal points is not
K r , j
empty: K r , j
0
Kr , j
r , j , 0 j m 1
r , j ,
K r , j K r , j
for various arbitrary Kr1, j , K r2, j
any K r , j r , j has a sectionally smooth regular boundary
1
2
Let us define the spaces of surface elements of the face K r : GrK, j
P
K r, j
L2 K r
r . j , 0 j m 1 .
Kr, j
K r , j
Similarly, by partitioning r , l , j , 0 j m 1 of a smooth patch K r , l of the face K r into the surface elements K r , l , j r , l , j , let us define the spaces of surface elements: GrK, l , j
P
K r, l, j
L2 K r, l
K r, l, j
r, l, j , 0 j m 1.
K r ,l , j
When constructing the external FE approximations of the Sobolev space H m , it is necessary to define m layers of surface elements on the boundaries of FE. It is possible that partitions of these layers do not coincide, i.e. if j
k then
r,l, j
r, l , k ,
0 j m 1,
r, j
r , k
0 k m 1
Let us define the following orthogonal projectors: RrK, j : L2 K r , j
PrK , j
(5.10)
BrK, j : L2 K r
GrK , j
(5.11)
RrK, l , j : L2 K r , l , j
PrK , l, j
(5.12)
BrK, l , j : L2
K r, l GrK , l, j
(5.13)
0 j m 1
Let us also define the orthogonal projector Bh : H 0
Gh
ˆ
that can be represented as follows (depending on the patches of boundaries being partitioned into the surface elements): Bh
m 1
B
K r, j
K
(5.14)
K r j 0
or: m 1
Bh
BrK , l , j . K
(5.15)
K r Kr ,l j 0
The orthogonal projector Bh is referred to as the operator of Gh -approximation. Let us call the components of the operator Bh defined by the expressions: m 1
B
K
BrK , j
(5.16)
Kr j 0
m 1
B K
BrK ,l , j K r Kr ,l j 0
as operators of G K -approximation.
From the definition of the operator of Gh -approximation and the operator of G K -approximation follows their important property that is established by the following theorem.
Theorem 5.1
Let us g be an arbitrary function from the space H 0 . Then ˆ
K
Bh g |K B K g .
(5.17)
The relation (5.17) is similar to the relation (4.9) between the operators of PK - and X h -approximation.
Approximations on a fixed partition Let us consider a patch
of the boundary of FE K . Let us assume that is an elementary
surface defined by the map : D
R n 1 R n
where is a homeomorphism; D is the domain of parameters t
t1,..., t n 1 of the map.
Since the partition of the domain is fixed, then the spaces H and H 0 are fixed as well. ˆ
2
The spaces L
are
ˆ
the elementary components of the spaces H and H 0 . Dense in the ˆ
ˆ
space L2 families of finite-dimensional spaces can be constructed in a number of ways. In particular, for this purpose, complete systems of algebraic polynomials and other systems of
analytical functions can be used. Below are closely examined the approximations of the spaces 2
L
by the spaces of the surface elements.
into the surface elements
Let us define a family of partitions h of the surface
K
h so that the partition requirements described in Section 5.2 are satisfied. Let us also
require that the family h is regular so:
there exists such positive constant that
K h hK K
(5.18)
where:
hK
diam K ,
K
sup diam S , S is a sphere from K
diameters hK approximate zero
The surface elements K
h are
scaled in the same way as FE K h in Section 4.3.1.
Let us designate for any surface element K
h through
K the element that satisfies the ˆ
following condition
diam(K )=1 . ˆ
Herewith K
where F is a similarity mapping:
F ( K ) , ˆ
(5.19)
F (t ) B t ˆ
ˆ
b,
t
ˆ
(t1 ,..., t n 1 ) , ˆ
ˆ
here B is a diagonal matrix of dimensions n 1 n 1 with diagonal elements equal to hK ; b is a vector of dimension n 1 .
Let us PK be a space of functions p : K
R determined on the domain K . Due to the
bijection of the mapping F one can associate with the space PK the following space
p : K R | p p F1 , p PK
P
ˆ
ˆ
ˆ
where P K is a space of functions determined on the domain K . Since F is a similarity ˆ
ˆ
mapping, then P
PK . Implication of the relations: t
ˆ
K t F t K ˆ
ˆ
p PK ˆ
ˆ
p p F1 PK ˆ
is the equality p t ˆ
ˆ
p t
t K , p PK . ˆ
ˆ
ˆ
ˆ
Theorem 5.2 Let us
K , P , K , P ˆ
ˆ
K
K
be two similar surface elements. Then, for the orthogonal
projectors: R : L2 K K
P
R : L2 K ˆ
holds the relation
K
PK
ˆ
ˆ
K
R v K
^
RK v ˆ
(5.20)
ˆ
v L K v v F
which is valid for all functions v L2 K , v L2 K that are in the relation ˆ
ˆ
2
ˆ
ˆ
◊ Proof –
Let us pi , pi , 1 ˆ
1
ˆ
L2 K
i N are the basic functions of the spaces PK and
K
P ˆ
correspondingly. Projection of the function v L2 K onto space PK is: N
K
R v
i v pi , i 1
where coefficients i v ,1 i
N are
determined from the following system of linear
algebraic equations N
v p , p i
i
j
i 1
0, K
v, p j 0, K ,
1 j N.
(5.21)
1 j N.
(5.22)
For the initial surface element K , the system (5.21) is: ˆ
N
v p , p ˆ
ˆ
i
ˆ
i
ˆ
i 1
j
0, K ˆ
v, p j 0, K , ˆ
ˆ
ˆ
Since F is the similarity mapping, then
p , p 0, i
j
K
pi p j dt
K
p p h ˆ
i
ˆ
j
n 1 K
dt ˆ
K ˆ
hKn 1 pi , p j 0, K ˆ
ˆ
ˆ
and the system (5.22) can be rewritten as follows
1 n
hK
N
v p , p ˆ
i
ˆ
i
j
i 1
0, K
hK1 n v, p j 0, K ,
1 j N.
(5.23)
The comparison of the systems (5.21) and (5.23) shows that
i v ˆ
ˆ
i v , 1 i N
and the required result is obtained.
Theorem 5.3
Let us K , PK be a surface element. If for some whole integer s 0 , the following inclusions ˆ
ˆ
hold
P
Ps K ˆ
ˆ
K
L2 K
then, there exists such a constant C , that for all surface elements
g
(5.24)
ˆ
K , P and any function K
H s 1 K the following inequality holds g
RK g l , K C hKs 1 K l g s 1, K ,
0 l s,
(5.25)
K K where R g is the projection of the function g onto the space P , and
◊ Proof
hK
diam K
K
sup diam S , S is a sphere from K
- On the space P K , the orthogonal projector R K coincides with the identical operator, ˆ
ˆ
i.e.
p PK ˆ
ˆ
K
R p ˆ
ˆ
p ˆ
Therefore, according to the inclusions (5.24)
p Ps K ˆ
ˆ
RK p ˆ
ˆ
p.
(5.26)
ˆ
The required inequality (5.25) follows from (5.26) and the relation (5.20) between orthogonal projectors RK , RK , and from the Theorem 3.4.1 from [10]. ˆ
Theor em 5.4 Let us there be a regular family of surface elements
K , P and for any element from the K
family the following inclusions hold
Ps K
PK L2 K , s 0 .
Then, there exists such a constant C that for all elements from the family and for any function
g
H s 1 K the following inequality holds g
◊ Proof
RK g
l , K
ChK s 1 l g s 1, K ,
0 l s.
(5.27)
- For every element K there can be put in correspondence some initial element K ˆ
K which satisfies the relation (5.19). Herewith the space P can be characterized as follows
PK
p : K R |
p
p F1 , p PK . ˆ
ˆ
ˆ
Then, for any element K from the family, the inequality (5.25) holds. The inequality (5.27) follows from the estimates (5.25), the conditions of regularity (5.18), and the bounded constants in the estimates.
Theor em 5.5 Let us h be a regular family of the partitions partitions of the surface patch into the surface elements K
K
h .
Let us assume that for some whole number s 0 and for any
h Ps K
PK L2 K ,
where Ps K is a space of polynomials of s -order. Then there exists such independent of h constant C that
g H s 1
BK h g
g
g L2
0,
Chs 1 g s 1,
lim g BK h g h0
0,
0
(5.28)
where
BK h : L2
GK h
is the orthogonal projector from the space L2 onto the space of surface elements: GK h
PK L2 ;
K
h
K
and:
h
◊ Proof
max hK , K
- From the Theorem 5.4, for any K
hK
diam K .
h the following inequality holds
(5.29)
g H s 1 K
g
RK h g
0, K
ChK s 1 g s 1, K .
According to the representation (5.29) of the space GK h , the following relation exists between the orthogonal projectors RK and BK h :
B h g |
g L2 , K h
K
K
RK g .
Taking into account the inequality
hK
h,
K
h
one obtains:
g H s 1 K g R g K
2 0, K
1
2
g B h g K
0,
2 Ch s 1 g s 1, K K
1
2
Ch s 1 g
s 1,
Let us s 0 then
g H 1
g
BK h g
0,
Ch g 1,
(5.30)
and
lim g BK h g h 0
0,
0.
The relation (5.28) follows from the relation (5.30) and the density of the enclosure
H 1
L2 .
To approximate the spaces H m , it is necessary to define m layers of the surface elements on the boundaries of FE. The parameter of a regular family of partitions of the face
Kr M K r (or smooth patch K r , l M K r , l of the face K r ) in general depends on the patch under consideration and on the layer number. Let us designate through r , j hr , j the family of partitions of the face K r for the j -layer of surface elements. Herewith hr , j
ma max x diam diam K r , j . K r , j
Similarly through r , l , j hr , l , j , let us designate the family of partitions of a smooth patch K r , l of face K r . Here hr , l , j
max x diam diam K r , l , j . ma K r ,l , j
The parameter h of the family of boundary approximations is defined by one of the following relations: h where K
,
Kr M K r ,
ma max x hr , j , K , Kr , j
(5.31)
0 j m 1,
or: h
where K
max
K , Kr , K r ,l , j
h , , , r l j
, K r M K r , K r , l M K r , l , 0 j m 1 .
Theor em 5.6
(5.32)
Let us be a fixed partition of the domain into FE K . Herewith each face K r of any FE
is
an
r , j hr , j , 0
elementary
surface.
Let
us
assume
that
there
are
regular
families
j m 1 of partitions of faces K r of every FE into the surface elements
Also, let us assume that for any surface element K r , j P0 K r , j
K r , j .
r , j hr , j the following enclosures hold
PrK , j L2 K r , j ,
0 j m 1
where PrK , j is the space of approximating functions of the surface element. Then, the family of spaces Gh of boundary functions defined by relations (5.7) and (5.8) is dense in the space H 0 , ˆ
i.e.
g H0
lim g Bh g H 0 ,
ˆ
h0
(5.33)
T
where:
Bh is the operator of Gh -approximation that is defined by (5.14)
parameter h is defined by (5.31)
◊ Proof
s
- In the Theorem 5.5 let us assume:
0, K r , h r , j hr , j , PK PrK, j , BK h BrK, j h r, j , 0 j m 1 ,
where operators BrK , j hr , j are defined by (5.11). Then for any K
g L2 K r
lim g BrK , j hr , j g
hr , j
0
, Kr M Kr 0, K r
0,
and the relation (5.33) follows from (5.31), (5.14), (5.16) and from the relation (5.17) between the operator of G K - and Gh -approximation.
Theorem 5.7 Let us be a fixed partition of the domain into FE K and let us every patch K r , l of any face of FE be an elementary surface. Let us assume that there are regular families r , l , j hr , l , j , 0
j m 1 of partitions of patches K r , l into the surface elements
let us assume that for any K r , l , j
r , l , j hr , l , j the following enclosures hold Ps j K r , l , j
PrK , l , j L2 K r, l, j ,
PrK ,l , j are the spaces of approximating functions of the surface elements, and
where s j
K r , l , j . Also,
0, 0 j m 1 .
Then, the family of spaces Gh of boundary functions defined by (5.7) is dense in the space H 0 , ˆ
i.e.
g H 0
lim g Bh g H 0 ,
ˆ
h 0
(5.34)
T
where the operator Bh and the parameter h are defined by the expressions (5.15) and (5.32) correspondingly. Further, there exist such independent of the parameter h constants C j , 0
j m 1 that
g W0
g
Bh g
m 1
C j h s
j
HT
j 0
1
g j,
where the space W 0 is defined by the characterization (5.4); j are semi-norms (5.5)
◊ Proof
- Let us assume that in Theorem 5.5:
(5.35)
s
s j , K r , l , h r , l , j hr , l , j , PK PrK , l , j , K
B
h
GK h
Br , l , j hr , l , j , K
GrK, l , j PrK , l , j , K r ,l , j
where s j
operators
BrK , l , j hr , l , j are
defined
by
the
expression
(5.13).
Let
us
also
0, 0 j m 1 . Then, from the Theorem 5.5 it follows that for any K
, K r M K r , K r , l M K r , l
g L2 K r , l
lim
hr ,l , j
0
g
BrK , l , j hr , l, j g
0, K r , l
0.
(5.36)
The relation (5.34) follows from:
the equality
g
Bh g
H T
m 1 g BrK , l , j hr , l , j g j 0 K K K r
the relations (5.36)
the definition (5.32) of the parameter h
the enclosure P0 K r , l , j
r ,l
P K , , , k
r l j
k
1
2 0, K r ,l
2
(5.37)
0
Let us s j 0, 0 j m 1 . According to Theorem 5.5, for any smooth patch K r , l of boundaries of FE K the following inequality holds
g H s
j
1
K , r l
g
BrK , l , j hr , l, j g
Chrs, l, 1j g j
0, K r ,l
s j
1, K r ,l
.
(5.38)
The required inequality (5.35) is the implication of (5.38), (5.37), and the definition (5.32) of the parameter h .
Approximations on a family of partitions Let us define a regular family of partitions h of the domain into FE K h . The parameter h of the family has a meaning of the maximum diameter of FE. When refining the partition, then
h 0 and the number of finite elements tends to infinity: card h
. In this case there are
the
the
families
of
the
boundary
spaces
considered
in
previous
sections:
HTh , H h , H 0h , Wh , Wh , W 0h . ˆ
ˆ
For any FE K h let us designate through K a similar FE that satisfies the condition ˆ
diam K 1 .
(5.39)
ˆ
Herewith:
diam K hK F : x R n F x Bx b R n K
F K ˆ
ˆ
ˆ
ˆ
where:
F is a similarity mapping
B is a diagonal matrix of dimensions n n with diagonal elements equal to hK
b is a vector of dimension n
(5.40)
Let us the patch
of the surface of the element K be defined by the map ˆ
ˆ
: D ˆ
ˆ
R n 1 R n ,
(5.41)
ˆ
where is a homeomorphism; D is the domain of the parameters t ˆ
ˆ
ˆ
t1 ,..., t n 1 of the map. ˆ
ˆ
Similarity mapping (5.40) transforms the coordinates of the point A in the global and local ˆ
ˆ
coordinate systems according to the following formulas:
hK xi bi , 1 i n 1 j n 1 t j hK t j ,
xi
ˆ
ˆ
Rn 1 of the map (5.41) is transformed into the similar
In this case the domain of parameters D ˆ
domain D R n 1 . Therefore, with the similarity mapping (5.40) is associated the similarity mapping F : t
ˆ
Rn 1 F t Bt Rn 1 , ˆ
(5.42)
ˆ
where B is a diagonal matrix of dimensions n 1 n 1 with the diagonal elements equal hK . Herewith D
F D . ˆ
Let us establish the relations between the points
g ˆ
t
ˆ
D, t D and the functions ˆ
L2 D , g L2 D as follows: ˆ
t
ˆ
D t F t D ˆ
(5.43)
ˆ
g : D R g g F ˆ
ˆ
The equality
ˆ
1
: D R (5.44)
g t ˆ
ˆ
g t
holds for all points t and t in the correspondence (5.43) and for any functions g and g in the ˆ
ˆ
ˆ
relation (5.44).
K Let us P be an enclosed space of boundary functions g :
R determined on the patch .
The mapping (5.42) associated with the mapping (5.40) establishes a one-to-one correspondence between points of domains D and D of patches ˆ
and for similar elements ˆ
K and K . Then ˆ
K with the space P there can be associated the space
P
g : R |
g
g F1 , g PK ˆ
ˆ
ˆ
that, due to linearity of the mapping F , coincides with space PK , i.e. P
PK .
Theorem 5.8 Let us P K and PK be the finite-dimensional spaces of functions defined on the boundary patches ˆ
and of similar FE K and ˆ
ˆ
K respectively. Then the orthogonal projectors:
R : L2 P K
K
2 R : L P ˆ
K
ˆ
ˆ
K
are in the relation
R g K
for all the functions g ˆ
^
RK g ˆ
L2 , g L2 from (5.44). ˆ
ˆ
(5.45)
◊ Proof -
The similarity mapping (5.40) of the element K is associated with the similarity ˆ
mapping (5.42). Scales of these mappings are identical. Further course of proof of this theorem is identical to the proof of the Theorem 5.2.
Theorem 5.9
Let us K be a finite element, D be the domain of parameters of the map of a boundary patch ˆ
ˆ
ˆ
of K , and P K be a finite-dimensional space of functions determined on D . ˆ
ˆ
ˆ
If for some whole number s 0 , the following enclosures hold
P
Ps D ˆ
ˆ
K
L2 D
(5.46)
ˆ
then, there exists such a constant C that for all elements K similar to K and for any function ˆ
g
H s 1 D g
RK g l , D C hKs 1 Dl g s 1,D ,
0 l s,
(5.47)
where: K
K
R g is the projection of the function g onto the space P
hK
diam K
D
sup diam S , S is a sphere from D
◊ Proof
- To the similarity mapping F : K ˆ
K it can be put in correspondence the similarity
mapping F : D D which has the same as F scale hK . On the space P K the orthogonal ˆ
projector
ˆ
P
RK : L2 D ˆ
ˆ
ˆ
K
is an identical operator. Herewith, according to the enclosures (5.46)
p Ps D ˆ
ˆ
RK p ˆ
ˆ
p. ˆ
Then the inequality (5.47) follows from the relation (5.45) between orthogonal projectors R K ˆ
and RK and from the Theorem 3.1.4 from [10].
Theorem 5.10 Let us there be given a family of finite elements, and any of the elements has a surface patch
K
with a map : DK
R n 1 K R n .
Let us assume that the family of the domains DK of the patches is regular, i.e. there exists such a constant that
DK
diam DK / D , K
(5.48)
where
DK
sup diam S , S is a sphere from DK .
If, for any FE from the family and for some whole number s 0 the following enclosures hold
Ps DK
PK L2 DK ,
where PK is a finite-dimensional space of functions determined on the domain DK , then there exists such a constant C , that for all elements from the family and for any function
g
H s 1 DK g
◊ Proof
RK g
l , DK
ChK s 1l g s 1,D , K
0 l s
(5.49)
- To any element from the family, it can be put in correspondence some similar element
K defined by the relations (5.39) and (5.40). Therefore, for every FE there exists such a constant ˆ
C (see Theorem 5.9) that
g H s 1 DK
g
RK g
l , DK
C hKs 1 Dl
K
g s 1, D .
(5.50)
K
The inequality (5.49) follows from the inequalities (5.50), the regularity conditions (5.48), and the bounded constants in the estimates (5.50).
Theorem 5.11 Let us define a regular family h of partitions of the domain into FE K h in such a way that every face K r of any FE from the family is an elementary surface. Let us assume that for every K h there are regular partitions r , j , 0 j m 1 of faces K r into the surface elements K r , j r , j . Let us assume that for any surface element K r , j r , j , the spaces PrK , j of approximating functions contain polynomials of zero degree: P0 K r , j
Then:
PrK , j L2 K r , j .
gh H 0h
lim gh Bh gh
ˆ
h0
H Th
where Bh is the operator of Gh -approximation defined by (5.14); h
◊ Proof
- By assuming in Theorem 5.10 that DK
0,
(5.51)
max diam K .
Kr , j , PK PrK, j , RK RrK, j , s 0
for any: K
h , K r , j r , j , 0 j m 1 ,
one obtaines:
g H 1 K r , j
g
RrK , j g
0, K r , j
ChK g 1, K , r, j
(5.52)
where RrK , j is the orthogonal projector (5.10).
Let us define the spaces: m 1
Z h
H 1 K r , K
Z 0h
K
h , K r M K r
Kr j 0
f Z h | r M r
fr1, j
1 fr2, j , 0 j m 1 j
Taking into account the estimate (5.52), the definitions (5.14) and (5.16) of the orthogonal projectors Bh and B K , and the inequality hK h , one obtains that for any function g h Z 0h m 1 g h RrK , j g h K K j 0 K r
2 0, K r , j
r,j
g h Bh g h
HTh
1
2
m 1 2 Ch g h 1, K K K j 0 K
r, j
r
Therefore
r, j
1
2
(5.53)
g h Z 0h
lim g h Bh g h h 0
0.
H Th
Then the relation (5.51) follows from the inequality (5.53) and the density of enclosure
Z 0h
H 0h . ˆ
Theorem 5.12 Let us define a regular family h of partitions of the domain into FE K h in such a way that each smooth patch K r , l of the face K r of FE K h is an elementary surface. Let us assume that for each K h are defined the regular partitions r , l , j , 0 j m 1 of the smooth patches K r , l into the surface elements K r , l , j r , l , j . Let us allow that for any surface element K r , l , j the spaces Pr ,l , j of approximating functions contain polynomials of s j -order s j K
Ps j K r , l , j
0 :
PrK , l , j L2 K r, l, j
Then there exist such constants C j , 0 j m 1 independent of the parameter: h
max hK , K
K h
that
g h W0 h
gh
gh H 0h
lim gh Bh gh
ˆ
h0
Bh g h
m 1
C j h s
j
HTh
j 0
H Th
0
1
gh
j
(5.54)
(5.55)
where Bh is the operator of Gh -approximation defined by (5.15); j is the semi-norm (5.5)
◊ Proof
- By assuming in the Theorem 5.10 that:
DK
Kr , l , j , PK PrK, l , j , RK RrK, l , j , s s j
one obtains
g H s
j
1
K , , r l j
g
RrK , l , j g
ChKs 1 g s j
0, K r ,l , j
j
1, K r ,l , j
,
where RrK ,l , j is the orthogonal projector (5.12). From this, considering the definitions (5.7) and (5.9) of the orthogonal projectors Bh and B K and the inequality hK h , it is obtained that for any function g h W 0h
gh
Bh g h
HTh
m 1 g h RrK , l , j g h K K K j 0 K r
r ,l
r ,l ,l
2 0, K r ,l , j
1
2
m 1
C h j
j0
s j
1
gh j .
The relation (5.55) follows from the inequality (5.54) and the density of the enclosure
W0 h
H 0h . ˆ
6. External approximations of variational equations Construction and total error estimate Let us be a domain in the space R n with a sectionally smooth regular boundary . Let us assume that V
H m , V0 H 0m
and let us consider the continuous on V
V bilinear
form a u, v
m
a
pq
p , q
where a pq x L
x D p uD q vdx ,
0
.
Generated by the bilinear form formal operator is defined by the relations:
u, v a u, v v V0 , u
m
p , q
1
q
D
0
q
a x D u . p
pq
The differential operator has order 2m . The space V
u
u
H m , u L2
is the domain of determination of the operator .
Elliptic boundary value problems for the operator are equivalent to the variational equations defined on enclosed subspaces W from H m such that H0m
W H m . The proof
of this fact is based on the existence of Green formula that corresponds to the given boundary
value problem. For instance, for a boundary value problem with mixed boundary conditions the space W can be defined by the characterization W
u H m | u 0 0
where is the trace operator (1.7); 0 is a bounded open sub-domain of the boundary .
The equivalency means that if the function u V satisfies some boundary conditions and the differential equation
u f ,
f
L2
(6.1)
then the function satisfies the variational equation as well
a u, v
f v v W
(6.2)
where f v is some linear form on W .
The external approximations of the equation (6.2) are constructed in the following way. Let us conduct the partition of the domain into FE K and let us define the spaces V , V0 , H , T by the relations (1.1) and (1.6). Let us introduce the continuous on V
V bilinear
form a u, v
K
where u
K
u
K
, vK v
a u , v onto the space V
K
m
a
pq
p , q
x D p u K D q v K dx
(6.3)
0 K
. It is apparent that the form a u, v is the restriction of the form
V .
The space X h V of FE approximates the space V . Discrete variational problem consists in the determination of the element uh
Vh R N h that satisfies discrete variational equation
vh Vh
f phvh
a ph uh , phvh
,
(6.4)
that is the external approximation of the equation (6.2).
Theorem 6.1 Let us the element f
V in the equations (6.2) and (6.4) is given and the following conditions
hold:
the form a u, v is W -elliptic, i.e.
C1 0 v W
a v, v
C1 v
a v, v
C2 v
2 V
(6.5)
the form a u , v is V -elliptic, i.e.
C2 0 v V
2 V
(6.6)
Then the variational problems (6.2) and (6.4) have unique solutions u -and uh correspondinglyand the following inequality holds
u
ph uh
V
C inf u phu h v V h
h
V
sup
wh V h
f ph wh ph wh V
a u, ph wh
(6.7)
◊ Proof – The
existence and the uniqueness of the solutions of problems (6.2) and (6.4) follows
from: the inequalities (6.5) and (6.6); the enclosure ph vh X h V for any h and for any vh
V h ; the Theorem 1.1.3 from [10]
From V -ellipticity and the continuity of the bilinear form a u , v , it follows: C2 ph uh
ph vh
2 V
a phuh ph vh , phu h phvh
a u ph vh , ph uh ph vh f p h uh ph vh a u , phu h ph vh Then
C2 ph uh
ph vh
V
C3 u ph vh
V
sup
wh V h
f
ph wh a u, ph wh ph wh
.
V
From this, taking into account the triangle inequality
u
phuh
V
u p hvh
V
phuh phvh
V
it is obtained the estimate (6.7).
The first term in the right-hand side of the estimate (6.7) satisfies the following inequality
inf u ph vh
vh V h
V
u Ah u
V
u AK u K
1
2 , m , K 2
(6.8)
where Ah is the operator of X h -approximation and AK is the operator of P K -approximation determined in Section 4.1. Then one can see that the first term depends on the smoothness of the exact solution u of the variational problem (6.2) and on the properties of the approximating K functions p P of the finite elements. The magnitude of this term is referred to as an internal
error of approximation. For the approximations built on a family of partitions of the domain and conformed by subspace FE, the estimates of the internal errors were obtained in Section 4.3.
Let us call as functional of the external error of approximation the magnitude
dh
a u , ph wh f ph wh .
The norm of the functional is referred to as an external error of approximation
d h
V
a u, ph wh
sup
f ph wh
ph wh
wh V h
.
(6.9)
V
External error estimates and convergence of approximations In order to obtain the estimate of the external errors let us derive the corresponding Green’s formula for a partition of the domain.
Generated by the form (6.3) formal operator is defined by the following relations
u , v a u , v u
K
u K ,
v V0 , K ,
K
u K
K
m
p , q 0
1
q
D q a pq x D p u K .
Let us be the trace operator (1.7) for the partition of the domain. For the spaces V , H , V0 , T and for the operator hold the assumptions of the Theorem 6.2.1 from [33] and therefore, there exists a unique operator such that it holds the following Green’s formula:
u , v V where
a u, v
u , v u , v ,
(6.10)
, is a duality relation on T T and where:
K
rK, l , j m
m
Kr Kr ,l j 2 m 1
rK , l , j ,
L H m K , H
m j 1
(6.11)
2
K , r l
(6.12)
j 2m 1
Theorem 6.2 Let us the variational problem (6.2) is approximated with FE that are built by the necessary and sufficient sign of the external approximations established in the Theorem 2.1 (i.e. conformed by subspace, homogeneous, or composite FE). Then there exists such independent of parameter h constant C that the external error of the approximation satisfies the inequality
dh
V
C u Bh u
H T
,
where:
H T is the space (2.2)
Bh : g
Gh is a space of the boundary functions
H T Bh g Gh is the operator of
Gh -approximation
(6.13)
◊ Proof
- From the relation (6.10) it follows that a u, ph wh
u, ph wh u, p h wh .
Let us multiply the equation (6.1) by the function ph wh and integrate it over the domain , so
u, phwh f , phwh . The following equality holds
u, ph wh u , ph wh , and then a u, ph wh
Let us identify the duality relation
f ph wh u , ph wh .
, on T T with the unique expansion of the inner product
, in H T . According to the Theorem 2.1 g h Gh ph wh X h
g h , ph wh H 0 , T
where X h is a space of FE. Then:
u V , g h Gh a u, ph wh
f ph wh u , ph wh H u g h , ph w h H T
u Bh u, p h wh H u Bh u T
H T
(6.14)
T
ph w h
H T
Here, the operator is continuous and it maps the space V into the space H T . Therefore, there exists such a constant C that
v V and then
v
H T
C v
V
,
ph wh X h V
ph wh
HT
C ph wh V .
(6.15)
The inequality (6.13) follows from the inequalities (6.14) and (6.15).
The external error is caused by two reasons:
by the discontinuity of the approximating functions from the space of finite elements on the inter-element boundaries
by approximate fulfillment of the essential boundary conditions of the boundary value problem
From the Theorem 6.2 it follows that, when the approximations are built by the necessary and sufficient sign, then the external error depends on the smoothness of the element u H T and on the approximating qualities of the spaces Gh of boundary functions.
Let us consider the approximations for which the spaces Gh are the spaces of surface elements introduced in the Section 5. There were formulated two approaches to define the layers of the surface elements:
in the first one it is assumed that the faces K r of FE undergo the partition into the surface elements (see Theorems 5.6 and 5.11)
in the second one it is assumed that smooth patches K r , l of faces K r of FE undergo the partition into the surface elements (see Theorems 5.7 and 5.12)
By assumption, each patch K r , l is an infinitely differentiable manifold in R n of dimension ( n 1 ).
Let us the solution u of the variational problem (6.2) belongs to the space
H 2m
k
,
m
0, k 0 .
Boundary operator defined by (6.11) and (6.12) maps the space H 2m
k
onto the product
of spaces m
H K
2 m k j 1 2
Kr Kr ,l j 2 m 1
K , (6.16) r l
Then, independently on smoothness of the exact solution u of the variational problem, the functions: u j*
r ,l , j
u,
m
j 2m 1
K r , l 2
belong only to the space L
K r even if their components are smooth enough: r , l , j u H 2 m k j 1 2
Functions u j* , m
K r, l .
j 2m 1are approximated by the surface elements from the space
(6.17)
Gh of
the boundary functions. If faces K r undergo the partition into the surface elements, then due to the insufficient smoothness of the function u j* , the estimates (6.13) will not depend on the smoothness of the solution u of the variational problem. Different situation occurs when into the surface elements are partitioned smooth patches K r , l of faces. In this case the surface elements approximate each component (6.17) of the product (6.16) that allows obtaining the estimates of
the external errors depending on the smoothness of the exact solution u and on the type of the boundary functions.
Theorem 6.3 Let
H 2m
us k
the
,
solution
m
u of
0, k 0 .
the
variational
problem
(6.2)
belongs
to
the
space
Let us assume that the external approximations are built with the
spaces Gh of the boundary functions that meet the conditions of Theorem 5.7 (or Theorem 5.12). Then there exist such independent of the parameter h constants C j , 0 j m 1 that m 1
dh
C j h s 1 u j , j
V !
(6.18)
j 0
where s j k j 1 , and where j are the semi-norms (5.3).
◊ Proof - For the components of the operator it holds the enclosure
rK, l , 2m j 1
L H 2m k K , H k j Kr ,l ,
0 j m 1.
Let us s j k j 1 then
L H 2m k , W 0 ,
where the space W 0 is defined by the characterization (5.4). Then inequality (6.18) follows from
Theorem 5.7 (or Theorem 5.12).
Theor em 6.4 If the approximations of the variational problem (6.2) are built on FE that satisfy the sufficient sign of the external approximations defined in Theorem 2.2, then the external error satisfies the inequality
dh
V
u
H T
sup
ph wh
ph vh
ph wh
wh V h
H T
,
(6.19)
V
where functions ph vh are defined according to the assumptions of the Theorem 2.2, i.e.
phvh
Yh H m
(6.20)
◊ Proof - From the enclosure (6.20) and Lemma 1.1, one has
ph vh Yh
u, p v
h h H T
0.
Then the required result follows from the inequality
a u, ph wh
f ph wh u, ph wh ph vh H u T
H T
ph wh
ph v h H
.
T
From the inequality (6.19) it follows that, if the approximations are built by the sufficient sign, then the external error depends on the approximating quality of the spaces Yh
H m of the
conformed approximating functions. Similarly one can obtain the estimate of the external error for the approximations built by the sufficient sign established in the T heorem 2.3.
Theor em 6.5 Let us the conditions of the Theorem 6.1 hold. Let us the FE approximations of the space V are the converging external approximations of the space W . Then
lim u phuh h 0
V
0
where u and uh are the solutions of the variational problems (6.2) and (6.4) correspondingly.
◊ Proof - When h 0 , the internal and the external errors in the right-hand side of the inequality (6.7) tend to zero, since by the assumption the approximations are external and converging in the space W .
Convergence of approximations in case when bilinear form is not V -elliptic Let us analyze the solvability and the convergence of the approximate solutions of a discrete variational problem in case when the solution of the continuous variational problem (6.2) exists and is unique, but the expansion a u , v of the bilinear form a u, v onto the space V
V
does not satisfy the condition (6.6) of V -ellipticity.
Theor em 6.6 Let us a partition of the domain into FE K be given and the following conditions hold:
A finite-dimensional subspace G ˆ
boundary functions grK , j , t , 0 ˆ
H 0 (see characterization (2.5)) is generated by such ˆ
j m 1 that
for any face
Kr M K r of
any FE
K the continuous on H m K functionals ir u
K r
K
K
g r , j , t r , j ud , ˆ
1 i M
(6.21)
are linearly independent on the space Pm 1 K of polynomials of m 1 -order
The enclosed subspace W W
V is defined by the characterization
g
u V | r M r
ˆ
1 1 1 r , j, t r, ju d
r
g
ˆ
2 r , j, t
r2, ju 2 d ;
r
g
K , K r M K r M K r
ˆ
K r, j, t
K r, j
(6.22) u d 0 K
K r
where and M K r are some not empty multitudes.
Then the mapping 1
u
W u
is a norm equivalent to the norm u
◊ Proof
V
W
2 2 u K m, K K
(6.23)
.
- On each FE K , let us define the projectors
rK : u H m K rK u Pm 1 K of the space H
m
K onto the space
Pm 1 K by the formula M
rK u ir u pir ,
(6.24)
i 1
where
M
p 1 is r i
i
some basis of the space Pm 1 K . Here the symbol r identifies the face
Kr M K r of
FE. The representations (6.24) exist because of the linear independence of
the functionals (6.21) on the space Pm 1 K . Let us
u
W
u
W
0.
Then on each FE
u K
u K Pm 1 K .
According to the characterization (6.22), for the projector that corresponds to the face
Kr M K r the following equality holds 1r u
K 1
0,
which implies
u
K 1
0.
(6.25)
1 2 Let us consider the element K that has some common with the element K face
K r2 K r1 r . From the equality (6.25) and the characterization (6.22) it follows that for this face
g
ˆ
r
1 r , j, t
1r , j u
K1
d g 2, , 2, u d ˆ
r j t
r j
K 2
0 .
r
2 It means that, for FE K , it can be pointed out such projector r 2 that
2r u
K 2
0,
and so
u
K 2
0.
Considering in a similar manner the rest of the finite elements of the partition, one comes to the conclusion that
K
u
K
0,
and so u
0 because each FE of the partition has at least one face that it shares with the
adjacent FE.
Theorem 6.7 Let us the assumptions of the Theorem 6.1 hold. Let us define the enclosed subspace W
V
with the norm (6.23) in accordance with the requirements of the Theorem 6.6. Let us the space X h
W of FE approximates the space W . Let us the form a u , v be W -elliptic, i.e.
C 0 v W
a v, v
C v
2 W
.
Then the variational problems (6.2) and (6.4) have the unique solutions u and uh correspondingly and the following inequality holds
u
◊ Proof
ph uh
W
C inf u phvh v V h
h
W
sup
f ph wh ph wh W
a u, ph wh
wh V h
(6.26)
- Proof of the Theorem is similar to the proof of the Theorem 6.1.
Theorem 6.8 Let us there be given a fixed partition of the domain . Let us the conditions of the Theorem 6.7 be satisfied. Let us the FE approximations of the space W be the converging external approximations of the space W . Then
lim u phuh h 0
V
0
(6.27)
where u and uh are the solutions of the variational problems (6.2) and (6.4) correspondingly.
◊ Proof
- Since approximations of the space W are the external approximations of the space W ,
then the internal error (6.8) and the external error (6.9) tend to zero when h 0 . Then relation (6.27) follows from inequality (6.26) by taking into account the equivalence of the norms
W
and V on the space W as established in the Theorem 6.6.
Theorem 6.9 Let us there be given a regular family h of partitions of the domain . For each partition from the family let us be satisfied the conditions of the Theorem 6.7. Let us FE approximations of the spaces W h , defined by the characterization (6.22), be the converging external approximations of the space W . Then
lim u ph uh h0
W h
0
(6.28)
where u and uh are the solutions of the variational problems (6.2) and (6.4) correspondingly.
◊
Proof – The relation (6.28) follows from the inequality (6.26) taking into consideration the
tendency towards zero of the internal and external errors of approximations.
The theorems show that even in the case when the bilinear form a u , v is not V -elliptic, it is possible to define some minimal set of boundary degrees of freedom generated by the space G ˆ
which provides the unique solvability of the discrete variational problem at any partition of the domain. If the approximations are built on a fixed partition of the domain, and if for any h , spaces Gh of the boundary functions contain the minimal subspace G defined in the Theorem ˆ
6.6, then takes place the convergence of the approximate solutions to the exact one, not only in the norm
W
, that is a semi-norm on the space V , but in the norm V as well. In general, for
approximations built on a family of partitions of the domain, the convergence can be established only in the norms
W . h
7. Examples of finite elements Justification of some classic non-conformed finite elements Many of classic non-conformed finite elements have been justified in works [9, 15 - 17, 22, 42 46]. Usually such justification is performed for a certain type of boundary value problem and the features of the differential equations of the problem are essentially used in the justification. Below, the theorems from the Section 2 are applied to prove that all of the examined FE generate converging external approximations of the corresponding spaces H
m
, m 1, 2 .
Spaces of the approximating functions of each of the elements under consideration contain the spaces of the polynomials of m -degree. This feature provides the convergence of the corresponding external approximations. Proofs of the convergence are not given here due to their simplicity.
Wilson's element
Wilson's element is used for the approximation of boundary value problems of the second order, whose generalized solutions belong to the space H 1 . The element is a parallelepiped in R n (Fig. 7.1.a.b) with vertices:
ai ,
1 i 4 n 1 .
The space P K of the approximating functions of FE can be defined by two equivalent ways:
n P P2 K L xi i 1 K P Q1 K L xi2 , 1 i n K
(7.1)
where:
L ... is the space generated by the functions listed in the brackets
P2 K is the space of polynomials of the second degree
Q1 K is the space of polynomials of degree, less or equal to one, on each of n variables, i.e.
p Q1 K
p
b
i 1 1 i n
1 ,... n
x1 1 x2 2 ...xn n .
The multitude of degrees of freedom of FE is defined as follows
p ai , 1 i 4 n 1 K 2 p . x 2 dx, 1 i n K i
Let us define a regular family h of partitions of the domain into FE. Let us introduce the space X h of FE, whose functions ph vh
X h are defined by their values in the vertices (nodes)
of the partition (here ph is an extension operator (1.4)) and b y their average values, as follows
2 ph vh |K K xi2 dx, 1 i n, K h
where ph vh |K P K .
The functions from the space X h are discontinuous on the inter-element boundaries and are continuous in the vertices of the partition. Therefore X h
H 1 .
Theorem 7.1 1
Wilson's element generates the external approximations of the space H
◊ Proof – Let
.
us consider an auxiliary finite element for which
P K
Q1 K
(7.2)
K p ai , 1 i 4 n 1 (7.3) Let us Yh C 0 be the space of the conformed FE defined by the relations (7.2) and (7.3), so
Yh
H 1 .
(7.4)
According to (7.1), the space Y h is a conformed subspace of the space X h .
For any function phuh
X h , let us designate through ph vh the unique function from the space
Y h that takes the same values as phu h in the vertices of the partition h . Let us estimate the
magnitude of phuh
phvh
H T
,
where space H T is defined by (2.2); is the trace operator (1.7).
One can write phuh
p hv h
K
Kr K r ,l
K r , l ,0
pK
rK, l ,0 p*K ,
where p K ph u h |K , p*K p h v h |K , and where rK , l ,0 p is a restriction of the function p P K onto the face K r K r , l of FE. Due to the construction of the function ph vh , the function rK, l ,0 p*K
P1 K r, l is a linear interpolant (when n 2 ), and a bilinear interpolant (when n 3
) of the function rK, l ,0 p K
P2 K r ,l . Then, according to the Theorem 3.1.4 from [10], there
exists such independent of the parameter h constant C that
rK, l ,0 p K
rK,l ,0 p*K
0, K r ,l
Ch2 rK,l ,0 p K
2, K r ,l
.
Therefore
lim phuh phvh h 0
H T
0
(7.5)
The inclusion (7.4) and the property (7.5) make the conditions of the Theorem 2.2 fulfilled.
Family of Crouzeix-Raviart ‘s elements Crouzeix-Raviart’s elements are intended to approximate generalized solutions of boundary 2 value problems of the second order. The domain K R is a simplex with the vertices
ai ,1
i 3 ; the space of the approximating functions is chosen in such a way that p P K
p |Kr ,l Pk
K r , l
(7.6)
where K r , l K r is a face of FE.
The multitude of degrees of freedom of the element is
K p bi , 1 i 3k , the points of Gauss’s integration quadrature of k -order (Fig. where bi K r , l , 1 i 3k are
7.1.c).
The space X h of Crouzeix-Raviart’s elements is not a subspace of the space H 1 .
Theorem 7.2 Finite elements from Crouzeix-Raviart’s family generate the external approximations of the space H 1 .
◊ Proof
1 2 - Let us consider the adjacent FE K and K . In nodes bi ,1 i
quadrature on the common face Kr1,l
k of Gauss's
K r2,l r ,l of adjacent FE, the approximating functions
from the space X h are continuous, i.e.
p1 P1 , p 2 P 2
p1 bi
p 2 bi 1 i k .
Since Gauss's quadrature of k -order is exact for polynomials of degree 2k 1 , and since the inclusions (7.6) hold, then the following relations hold too
p1 P1 , p 2 P 2 , g Pk 1 K r , l
g p1
r ,l
1
Kr ,l
d g p
2
r , l
K r2,l
(7.7)
d
Let us associate with the family of partitions h the family of spaces H 0h defined by (2.3), and ˆ
let us define the family of spaces of boundary functions: Gh
G
K r , l, 0
K
where GrK , l ,0
H 0h , ˆ
K r K r ,l
Pk 1 K r , l .
According to the Theorem 5.12, the family Gh is dense in the space H 0 where the space H 0 is ˆ
the limit space for the family H 0h . ˆ
1 2 For the adjacent FE K and K , functions from the spaces GrK ,l ,0 satisfy the relation
g1r , l ,0
gr2,l ,0 ,
gri ,l ,0
Gri,l ,0 , i 1, 2 .
Using the designations introduced, the relation (7.7) can be rewritten as:
g1r , l ,0 r1,l ,0 p1d
K 1r ,l
gr2,l ,0 r2,l ,0 p 2 d 0 ,
K r2,l
where rK ,l ,0 are the components of the trace operator (1.7).
Therefore, the following relation holds
g h Gh , ph uh X h where is the trace operator (1.7).
g h , ph uh H 0 , T
ˆ
Therefore, the conditions of the Theorem 2.1 are met and the approximations built on the elements of Crouzeix-Raviart are the external approx imations of the space H 1 .
Let us note that according to the Theorem 6.3, for sufficiently smooth solutions of variational problems, the external error of approximation has order O h k .
Morley's element Morley's element is used to approximate the solutions of boundary value problems of fourth order. The domain K
R2 is
a triangle with the vertices ai ,1
i 3 . The space of the
approximating functions of FE is:
P K
P2 K .
(7.8)
The multitude of degrees of freedom:
p ai , 1 i 3 K p aij , 1 i, j 3 n where
p aij n are the derivatives along the normal in the middles of the sides of the triangle (Fig. 7.1.d).
(7.9)
Let us define a regular family h of triangulations of the domain . The approximating functions from space X h of Morley's FE are continuous in the vertices of triangles K h . Derivatives along normal to the FE boundary are continuous in the middles of the sides of triangles only. Therefore, one has the inclusion X h
L2 instead of required X h H 2
.
Theorem 7.3 Morley's element generates the external approximations of the space H 2 .
◊ Proof
- Let us define a triangular FE K
R2 for which: P K
P1 K
K p ai , 1 i 3 .
(7.10) (7.11)
The corresponding space of FE is designated through Y h0 . The element defined by (7.10) and (7.11) is conformed, i.e. inclusion Yh0 C 0 holds, so
Yh0
H 1
(7.12)
Due to the characterizations (7.8) and (7.9) the inclusion Yh0 X h holds as well.
For each function phuh
X h , through ph vh let us designate a function from the space
0
Y h that
has the same values as phu h in the vertices of the triangulation h . Let us introduce the designations
p h u h |K , p * p h v h | K .
p
On each face K r , l of FE K h , the function rK , l ,0 p* is a linear interpolant of the function
rK , l , 0 p , so rK, l ,0 p
rK,l ,0 p*
0, K r ,l
Ch2 rK ,l ,0 p 2,K
r ,l
and, therefore
ph uh X h
lim 0 ph uh 0 ph vh h 0
H T
0,
(7.13)
where 0 is a trace operator of zero order for the partition of the domain. It is defined by the expression (2.16).
1 2 Let us consider the adjacent FE K and K . The derivatives along normal to the common side of
the adjacent FE are continuous in the middles of the sides, i.e.
ph uh |K n
ph uh X h
By designating through
1
aij
ph uh |K n
2
a . ij
r ,l Kr1,l K r2,l this common side, one can write:
phuh X h
r1, l ,1 p1
r2, l ,1 p 2 ,
(7.14)
where ri ,l ,1 is the operator of differentiation of the first order along outward normal to FE boundary; p
i
phuh |K , i 1, 2 . i
The middle of the common side of the adjacent FE is a point of the Gauss's integration quadrature of the first order that is exact for pol ynomials of the first degree. Since
p P K P2 K
rK , l ,1 p
P1 K r , l ,
r2, l ,1 p 2 d .
then from the relation (7.14) it follows that
r1, l ,1 p1d
1
(7.15)
K r2,l
K r ,l
Let us define the family of spaces of the boundary functions (see characterization (2.18)) 1
1 j GrK , l , j H 0,1 h ,
1
Gh
K
where GrK , l ,1
(7.16)
ˆ
Kr Kr ,l j 0
P0 K r , l .
According to the Theorem 5.12, the family Gh1 is dense in the space H 0,1 that is the limit space ˆ
for H 0,1 h when h 0 . ˆ
For the adjacent FE, due to the inclusion (7.16), the following equality holds:
g1r ,l ,1
gr2,l ,1 ,
gri ,l ,1
Gri,l ,1 , i 1, 2 .
Then (7.15) can be rewritten as follows
g 1r , l ,1 r1,l ,1 p1d
1
K r ,l
gr2,l ,1 r2,l ,1 p 2 d 0 .
(7.17)
K r2,l
and therefore
g1h Gh1 , ph uh X h
g1 , 1 p u h
h
h H T
0,
(7.18)
where 1 is the trace operator of the first order for the partition of the domain. The operator is defined by the expression (2.16).
From the inclusion (7.12), from the relations (7.13) and (7.18) and from the Theorem 2.3 follows that the approximations built on Morley's FE are the external approximations of the space
H 2 .
Fraeijs de Veubeke’s element Fraeijs de Veubeke’s element is used to approximate generalized solutions of the boundary value problems of the fourth order. The domain K R 2 is a triangle and the space of the approximating functions is:
P K where f
p is
p P3 K |
f
p 0 ,
some functional whose expression is not necessary here. The multitude of
degrees of freedom of FE is as follows
p 1 K p ai , 1 i 3; p aij , 1 i j 3; d , meas K r , l K n r ,l
where:
ai , aij are the vertices and the middles of sides of the triangle correspondingly
p n is the derivative along normal to the side K r , l K r of the triangle (Fig. 7.1.d)
Let us h be a regular family of partitions of the domain into FE K h . The approximating functions from the space X h of Fraeijs de Veubeke’s FE are continuous in the vertices and the middles of sides of the triangles K h only, i.e. the inclusion X h
L2 holds.
Theorem 7.4 Fraeijs de Veubeke’s element generates the external approximations of the space H 2 .
◊ Proof
- Proof of the Theorem is similar to the proof of the Theorem 7.3. The only differences
are that the conformed space Y h0 has to be built on the quadratic triangular FE defined as follows: P K
P2 K ,
K p ai , 1 i 3; p aij , 1 i j 3 and that the relation (7.17) immediately follows from the definition of the degrees of freedom of Fraeijs de Veubeke’s FE.
Elements for external approximations In the classic Finite Element Method, strict requirements are imposed on the element shapes, on the location of the element nodes, and on the approximating functions of the element. This is caused by the necessity to provide sufficient smoothness of the approximating functions from the space of FE on the inter-element boundaries. The only requirement to the geometry of FE for external approximations is that it must have sectionally smooth regular boundary. Such FE can be of arbitrary shape. Boundary degrees of freedom of FE are generated by boundary functions and they are defined independently on the types of approximating functions that represent field variables inside FE. This property is reflected in the definitions of P -multi-solvency of the multitude of boundary degrees of freedom of FE (see Section 3.1). Internal degrees of freedom of FE are defined automatically after boundary degrees of freedom and a space of approximating functions of the FE has been specified.
In this Section are given some examples of FE for the external approximations of 2D problems, axis-symmetrical problems, and 3D problems. Also, the ways of parameterization of boundaries of FE and the approaches to the definition of boundary degrees of freedom are discussed. Herewith it is assumed that spaces of approximating functions of FE are complete spaces of
Pk K .
polynomials of k -degree, i.e. PK
With such choice of spaces P K any function
p P can be expressed as follows K
p
k
b x
,
x K ,
0
where
1,..., n is a multi-index;
b are some coefficients.
The number k is called the order of internal approximation on FE. Polynomial approximations are convenient in computations and, as at will be demonstrated in Section 8, they provide high accuracy of solutions of boundary value problems.
It is assumed also that a face of the element consists of one smooth patch only, so K r K r , l . Preference for such approach is explained in Section 6.
2D elements for boundary value problems of second and fourth orders
A smooth patch K r , l of the 2D FE boundary is defined by the mapping
: D
R Kr ,l R 2 ,
(7.19)
where D is the domain of determination of the parameter t . Let us assume that the mapping (7.19) is a polynomial of
l -degree.
When l
1 , the patches K r , l are the curves of second,
third, and higher orders. In general, the parametric equations of a curve of l -order can be written in the form l 1
xi
f j t xi j ,
1 i 2,
(7.20)
j 1
j
where xi , 1 i
f j t , 1 j
2, 1 j l 1 are the coordinates of the nodal points of the curves;
l 1 are some polynomial functions.
Below there are the expressions for the functions when l
l
3, t D 1,1
1 (mapping 1 ): 1 2 1 t f 2 1 2 1 t f1
l 2 (mapping 2 ):
1 2 t t 1 f 2 1 2 t 1 t f1
f3
1 t 1 t
l 3 (mapping 3 ): f1 1 16 1 t 10 9 t 2 1 f 2 1 16 1 t 10 9 t 2 1
2 f 3 9 16 1 t 1 3t
f 4 9 16 1 t 2 1 3t
(see Fig. 7.2):
From practical point of view the expressions above are sufficient. More detailed description of the boundary of the domain is best done through increase the number of smooth patches that approximate the boundary instead of through the increase of the order l of the curves.
Parametric equations of smooth patches of FE boundary are completely defined by the coordinates of their nodal points. Let us recall, however, that in the suggested method, the degrees of freedom are not associated with the nodal points. Nodal points, in this case, are only necessary for defining the shape of the element.
Let us consider how to construct the spaces G K of the boundary functions of the element. Solutions of variational equations, which correspond to boundary value problems of 2m -order, belong to the space H m . Herewith the order of normal derivatives that appear in the definition of boundary degrees of freedom of FE changes in 0
j m 1 , and on the boundary
of FE it is necessary to define m layers of surface elements.
Let us r , l , j , 0 j m 1 be partitions of smooth patches K r , l of FE boundary into the surface elements K r , l , j r , l , j , 0 j m 1 . For different values of the index j the identical partitions can be used, i.e. r , l , j r , l , i , 0 i m 1, 0 j m 1 . On each K r , l , j it is K
necessary to define spaces Pr ,l , j of approximating functions. According to the Theorems 5.7 and 5.12, the sufficient condition of convergence of the approximations is the following inclusion P0 K r , l , j
PrK , l , j , 0 j m 1 .
Here and below Pi designates the space of polynomials of i -order defined on the corresponding domain. If the following inclusion holds Ps j K r , l , j
PrK , l , j , 0 j m 1
then, as it is shown in the Section 6, the external error of approximation of the variational equation with sufficiently smooth solution has an order O h s , where s
min s j 1 , 0 j m 1 . j
In the benchmark problems of the Section 8, the following space of polynomials of s j -order is used as the space of approximating functions of the surface elements K
Pr , l , j
Ps K r , l, j , 0 j m 1 . j
The number s j is called the order of the boundary approximation of j -layer of the surface elements. Herewith the space G K of the boundary approximating functions of the element is defined as follows m 1
G
K
GrK , l , j ,
(7.21)
Kr Kr ,l j 0
where GrK, l , j
P
K r, l, j
L2 K r, l
K r ,l , j
is the space of surface elements of j -layer.
Let us also point out one important class of the spaces G K characterized by that the patches
K r , l are partitioned into one surface element, i.e. Kr, l , j
D, card r , l , j 1, 0 j m 1 .
(7.22)
Herewith: GrK , l , j
Ps D , 0 j m 1 . j
Basic functions of the space G K generate boundary degrees of freedom of FE. According to the representation (7.21), the space G K is the union of the spaces: Ps j K r , l , j t
P t , sj
Kr, l , j , 0 j m 1
The simplest system of the basic functions of the space Ps j t is the system of monomials s j
t 0 . When the relations (7.22) hold, the functions i
i
f j ,1 i
l 1 from (7.20), that describe
the shape of FE, can be used as the basic functions of the spaces Ps j t as well.
Let us: s j
g t 0 i j
i
be some basis of the space Ps j t . Then the corresponding boundary degrees of freedom are defined as follows
p P K
ji
p
i
K
g j t r , l , j pd ,
(7.23)
K r ,l , j
0 j m 1, 0 i s j When j
0 , the function rK ,l ,0 p is the restriction of the function rK , l ,0 p p K . r ,l
When j
1 , then
p onto the patch K r , l :
rK , l ,1 p
p p n1 n , x1 x2 2
where n1 , n2 are cosines of the outward normal to the boundary of FE. The surface measure is: 1
dx1 2 dx2 2 2 d dt dt , dt where x1 t , x2 t are the parametric equations of the patch K r , l .
3D elements
Let us consider some examples of the polynomial mapping
: D
R2 Kr ,l R3 ,
that can be used to define a smooth patch of the boundary of 3D element.
In Fig. 7.3 are shown the patches of faces of FE whose parametric equations have the form xi
N
f
j
t xi j ,
1 i 3,
(7.24)
j 1
where t
t1, t 2 ;
j
f j t are some polynomials; xi are the coordinates of the nodal points of
the patches K r , l . Here are the expressions for the functions f j t ,1 j shown in Fig. 7.3:
K r , l is a flat triangle, N 3 (mapping 1 ):
N for the patches
t 1 f 2 t 2 f3 1 t1 t 2 f1
K r , l is a triangle of second order, described by the complete polynomial of second degree, N 6 (mapping 2 ): f1
t1 2t 1 1
f 2
t2 2t 2 1
1 t1 t2 1 2t1 2t 2 f 4 4t1t 2 f 5 4t 2 1 t1 t 2 f 3
f 6
4t1 1 t1 t 2
K r , l is a flat quadrilateral, N 4 (mapping 3 ): f1 1 4 1 t1 1 t 2 f 2 1 4 1 t1 1 t 2 f 3 1 4 1 t1 1 t 2 f1 1 4 1 t1 1 t 2
K r , l is a curvilinear surface of third order, N 8 (mapping 4 ):
1 4 1 t1 1 t 2 t1 t 2 1 f 2 1 4 1 t1 1 t 2 t1 t 2 1 f1
1 4 1 t1 1 t 2 t1 t 2 1 f 4 1 4 1 t1 1 t 2 t1 t 2 1 f 3
f 5
1 2 1 t12 1 t 2
f 6
1 2 1 t22 1 t 1
f 7
1 2 1 t12 1 t 2
f 8
1 2 1 t22 1 t 1
Domains D of the parameters of maps shown on Fig. 7.3 have simple shapes and that simplifies their partition into the surface elements K r , l , j r , l , j , 0 j m 1 .
Let us assume K r , l , j D, 0 j m 1 i.e. let us consider the whole domain D to be one surface element. In this case the system
t ,
t
s j ,
t1 , t 2 , 1 , 2
can be chosen as the system of the basic functions of the surface element K r , l , j . The system of functions f j t , 1
j N from
the parameterization (7.24) also can be used for the same
purpose.
Boundary degrees of freedom are defined, as in case of 2D FE, by (7.23). Herewith: d
1
g12 g22 g32 2 dt
where:
dt1dt2 x x x x g1 2 3 2 3 t1 t2 t2 t1 x x x x g2 3 1 1 3 t1 t2 t1 t 2 x x x x g3 1 2 2 1 t1 t2 t1 t 2
dt
Elements for approximations of spaces of vector functions
Let us assume that the solution of some boundary value problem is a vector function of variables x
n
xi i 1 R n , i.e. U x
n
ui x i 1 ,
and let us assume that each component ui ,1 i
(7.25)
n of the function belongs to the space
H m . Then the variational equation is defined on the space V
n
V , where V H m is
a space of vector functions. For example, when solving the problems of Theory of Elasticity in displacements, the function (7.25) is a displacement vector and V
n
H 1 .
In computations, each component of the space V is approximated independently of the others by using the FE described above. Therefore, for each FE are constructed the independent representations of the components N i
u
k
b x
i
,
1 i n.
0
There can be another choice of boundary degrees of freedom of FE different from that defined by (7.23). It can be caused by the following reasons. Often the boundary conditions are imposed not onto the components ui ,1 i
n of vector in global coordinate system, but onto the
components u i ,1 i n 1 of the vector in natural coordinate system, that has coordinate axes normal and tangent to the boundary of the domain. For example, when solving the problems of Structure Mechanics, one may need to fulfill the boundary condition when the normal to the boundary component of the displacement vector equals to zero: un
0 and the tangent
displacements u i can be arbitrary (slippery support). Similar situation occurs when it is necessary to take into account the cyclic symmetry of the structure. Besides, if the approximations of vector functions on adjacent FE are constructed in different coordinate systems, then when forming the system of equations it is necessary to bring the local boundary degrees of freedom of FE to the global ones.
One of the possible ways of effective solution of the above mentioned problem is the construction of vector approximations on FE. Here the displacement vector of FE is represented as follows N
U
N
bi i , i 1
where i ,1 i
N are some vector approximating functions.
Let us, for example, n 2 . Then the simplest choice of representation (7.26) is
U
N
1i x b x i 1 0 2i 2
k
i
where ij
0, i j 1, i j
Therefore, the space PK of approximating functions is a vector space P
K
Pi K , 1 i n
(7.26)
and, as the boundary degrees of freedom, it can be taken the integrals from the normal and tangent displacement components.
When n 2 , taking into account the relations (7.23), the boundary degrees of freedom can be written as follows
p p1 , p2 P K P1K , P2K 0i n p
g 0 t r , l , 0 p1 n1
rK, l, 0 p2 n2 d
g 0 t r , l , 0 p2 n1
rK, l, 0 p1 n 2 d
i
K
(7.27)
K r , l ,0
0i p
i
K
K r , l ,0
where n1 , n2 are the cosines of the outward normal to the boundary of FE. Elements with the boundary degrees of freedom of type (7.27) are referred to as vector FE. In the Section 8 the vector FE are used when solving the 2D and the axisymmetrical problems of the Theory of Elasticity.
Axisymmetrical elements 3 Let us assume that the domain R is a body of revolution. In this case, taking into account
the axial symmetry, the geometry of the domain can be described by its axial cut in the cylindrical coordinate system. If the boundary conditions of the boundary value problem and the right-hand side of the corresponding differential equation are such that the solution of the problem does not depend on the angle coordinate, then the problem is called axisymmetrical. From the mathematical point of view, axisymmetric problems are 2D problems because the
sought functions depend only on the coordinates x1 , x2 . Herewith the analysis domain is a half of the axial cut of a body and is located in the semi-plane x1 0 .
In axisymmetrical problems of Theory of Elasticity the displacement vector U
u1 , u2 has the
radial component u1 and the axial component u2 . On the axis of symmetry the radial component of displacements turns into zero, therefore the vector U belongs to the product of spaces V1
V 2
where: V1
u1 H 1 |
V2
H
u1
0 when x1 0
1
(7.28)
Let us partition the domain into FE. The representation of the displacement vector of FE can be in the form (7.26) and the boundary degrees of freedom in the form (7.27). Herewith in the relations (7.27) the surface measure is: 1
dx1 2 dx2 2 2 d 2 x1 dt dt dt
where x1 t , x2 t are the parametric equations of the patch K r , l .
Space of the approximating functions of FE can be of two kinds:
If FE does not touch the axis of symmetry OX 2 , then the space is: P
K
Pk K
2
If FE touches the axis of symmetry then, due to the characterization (7.28), the approximant of the radial component of the displacement vector has to satisfy the condition: u1 N
0
when
x1
0.
(7.29)
The condition (7.29) holds if it assumed that u1 N x1 Pk ( K ) . Therefore, for the FE touching the axis of symmetry OX 2 :
P K
Pk K x1Pk (K ) .
Finite elements for stress concentration In classic FEM, analysis of stress concentration around holes, notches, and other small details of structures, requires significant mesh refinement in domains with high stress gradients. Herewith, the dimension of the discrete problem increase, and the conditioning of the equation system worsens. Here, the approach developed in the previous sections is used to construct special conformed by subspace FE with displacement and stress approximation in complex variables. The FE is intended to simulate stress concentration domains. Unlike some other approaches [53 55], that also use local asymptotic representations of solutions of boundary value problems, in external approximations it is not necessary to construct any particular functionals that take into account the discontinuity of approximating functions on inter-element boundaries. Completeness of the system of complex basic functions in the domain of special FE and the fulfillment of the general criteria of the convergence of the external approximations provide the convergence of approximate solutions to the exact. This is supported by the results of the numerical experiments described in Section 8.
Let us consider a polygonal FE meant for the simulation of displacement and stress field in the vicinity of an elliptic hole or a notch. The FE domain can be either double connected, when the hole is located completely within the FE (Fig. 7.4.a), or single-connected when a part of the contour of the element is an elliptical arch (Fig. 7.4.b). Lengths of the semi-axis of the ellipse can be arbitrary, which in combination with the possibility of changing the FE connectivity, allows to handle a rather large class of stress concentration problems.
Let us use the methods of the theory of functions of complex variable z
x1 ix2 to obtain the
local asymptotic representation of displacements and stresses that exactly satisfies boundary conditions on the contour of the hole as well as the equilibrium equations. Using two analytical functions,
z and X z (complex potentials), displacements and stresses in the domain of
special FE can be represented as follows
2G u1 iu2 C1 z z z X z
11 22 2i 12 2 z z X z 11
where G
22 4 Re z
(7.30)
E 2 2 ; E is modulus of elasticity; is Poisson's ratio; C 1 3 1
for plane stresses; C 1
3 4 for plane strains; denotes differentiation on
z ;
denotes
complex conjugate.
Any FE with an elliptical hole can be obtained from some initial FE with a round hole by the conformal mapping
z
C m 1 ,
where:
i ; C
a
b 2
; m
b ab a
The reverse mapping:
z
z 4C m 2
2
1
2
2C
maps the element with an elliptical hole onto the element with a round hole (see Fig. 7.4.a). Conformal mapping allows to construct the approximation on the complex plane instead of z and, therefore, the problem of boundary conditions fulfillment at the contour of the elliptic hole is replaced by a simpler problem of doing the same thing for the unit round hole.
Let us substitute the variable in the equations (7.30 ):
z 1 X z
X X 1
Then:
1
X 1 X z 1 1 z 3 X 1 X 1 X z 3 z
(7.31)
Substituting expressions (7.31) into (7.30) and leaving out the index 1 one obtains:
2G u1 iu2 C1
X
11 22 4 Re 11 22 i 2 12 2 X X
Any function , analytical in some circular ring 1
3
(7.32)
R , can be uniquely represented in
this ring by Laurent expansion
a0 j j j j , j 1
where j 1 j i 2 j are complex coefficients. Let us approximate analytical functions
and X by the finite segments of Laurent
expansion
N 0 j j j j j 1 N 2 N j j X N 0 j j j 1 j 1 N
(7.33)
Let us name the number N here as the order of internal approximation on complex FE. Unlike the work [56], in (7.33) are used the segments of complete Laurent expansion and all the members, up to the particular degree, remain.
Let us assume the contour of the hole to be free from the load. Then at
ei the
following condition must hold:
X
. 0
(7.34)
Substituting the equation (7.33) into (7.34) and reducing at the same degrees of , the equations for the complex coefficients are obtained:
1 C m1 1 1 m1 2 C m 0 2 2m 2 j C m j 2 j j 2 j 2 mj j , 3 j N 2 j C m j 2 j j 2 j 2 mj j , 1 j N N 1 N 2 N 1 N 2 0 0 C 0 m 2 2 2
(7.35)
Making the reverse substitution of equations (7.35) into (7.33) it is obtained the following representations of the complex potentials:
j 1 N j 2 mj j X N j C j j 1 N j 2 j j C m j 0 N j C j j 2 mj j j 1 N j 2 j j C m j 1 N
N 0 j j j j
(7.36)
The complex coefficients in (7.36) are:
j1 i j 2 j j1 i j 2
j
where j1 , j2 , j1 , j 2 are real numbers. Then relations (7.36) can be rewritten in the form:
j 1 j j j1 j 2i 2 2 01C m 1 02iC 1 m X N N j1C j j 2 mj j j m j 2 j 1 j 2iC j j 2 mj j j m j 2 j1C m j 2 j j 2 mj j j j 2iC j j j 2 mj j m j 2 N
N 01 02i j1 j j 2i j
(7.37)
Let us denote the complex basic functions in (7.37) through j and j correspondingly, for 1 j through b j for 1 j
4 N 2 , and let us denote the real factors of these functions
4 N 2 . Let us introduce the vector potential: F
X
Then, the approximant F N of the vector potential F is:
N F N X N
4 N 2
b F , j
(7.38)
j
j 1
where:
j F j j are
vector
complex
basic
functions.
Basic
functions
4 N 2
F 1 j
j
generate
the
K finite-dimensional space of vector complex functions that approximate vector potential
F on FE.
Substituting (7.38) into (7.32) and taking into account that
1 z it
is obtained the
representations for the displacement vector of the element on the -plane and on the z -plane:
j0 bj 0 j j 1 u1 N z 4 N 2 j0 1 z U N z N bj 0 1 u2 z j 1 j z 4 N 2 j z 4 N 2 b j b jU j z j z j 1 j 1
u1 N U N N u2
4 N 2
(7.39)
ˆ
ˆ
where j0 , 0j , j , j are some complex basic functions. ˆ
ˆ
4 N 2
By the terminology of Section 3.1 the system U j z
j 1
is the initial system of functions
necessary to build basic functions of complex FE corresponding to its boundary and internal degrees of freedom. These functions generate the space P K of the approximating functions of FE. To construct the basic functions of FE it is necessary to determine the explicit form of functions U j z ,1 j
4N 2 . This is a rather difficult problem. However, there is a way to
calculate stiffness matrix of FE that does not require explicit expressions for the basic functions of FE.
Due to one-to-one correspondence of the elements from spaces K and P K , for each function
U j z ,1 j
4N 2 there is a function F j ,1 j 4N 2 whose explicit expression is
known. Therefore, instead of finding the representation U N z for the displacement vector of the element through its boundary and internal degrees of freedom, it is sufficient to find such representation for the vector potential F N . In accordance with the procedure described in Section 3.1, it is necessary first - to build the system (3.12) of equations for the factors
b j ,1 j
4 N 2 of the initial representation (7.38) of the vector potential, and second - to
construct the new representation of the potential by the formula (3.16).
Let us consider how the elements of the matrix of the system (3.12) are calculated in this case. Referring to the Section 7.2, boundary degrees of freedom of FE on z -plane are as follows: 0i n
g 0 t n1u1 i
N
z t n2u2 z t d N
K r , l ,0
0i
N N g 0 t n1u2 z t n2u1 z t d i
(7.40)
K r , l ,0
where:
x1 t ix2 t is a parametric equation of the boundary patch
z t
K r , l ,0 is the surface element from zero layer
g0i t is the basic function of the surface element
n1 , n2 are cosines of the outward normal to the boundary patch
From expressions (7.39) and (7.40) it follows that to calculate the matrix of the system (3.12) one needs to calculate the integrals: t2
g
i
0
t j z t d , 0
t1
t 2
g
i
0
t 0j z t d ,
t 1
that can be performed numerically (for example, by Gauss’s quadrature). To calculate values of the functions j z t and j z t in the nodes tr ,1 r ˆ
ˆ
M of the quadrature, it is not
necessary to know the explicit expressions for these functions. It can be done by the formula
j 1 C 1 j j z i j z 2G ˆ
ˆ
ˆ
ˆ
ˆ
j
ˆ
that is evaluated in the points of -plane: r
1 z t r .
Representations (7.30) for displacements and stresses satisfy the equilibrium equations in the domain of special FE when the mass forces are absent. In this case the corresponding expression for bilinear form in the variational equation can b e transformed by Green's formula 2
ij
K i , j 1
where u
u ij v dx
2
ij
u n j vi d
K i , j 1
u1 , u2 , v v1 , v2 .
Therefore, when calculating the stiffness matrix, integration is performed along the contour and not on the domain of special FE. It reduces significantly the amount of calculations.
Figure legends for Section 7:
Fig. 7.1 Examples of non-conformed FE
Fig. 7.2 Examples of 2D faces defined by polynomial mappings
Fig. 7.3 Examples of 3D faces defined by polynomial mappings
Fig. 7.4 Stress concentration element on z - and -plane: (a) element with elliptical hole (b) element with elliptical arch
8. Numerical experiments This Section presents the results of numerical experiments for various benchmark problems of Structure Mechanics. The accuracy of the suggested method of external FE approximations (MEFEA) is estimated by comparing the approximate solutions with the exact analytical solutions, as well as with numerical and experimental results obtained by other authors. The influence of various parameters of approximation on the accuracy is closely examined. It is demonstrated that results of the numerical experiments are completely in conformity with the theoretical estimates obtained in the previous sections.
All the experiments are performed using the FE described in Section 7.2. Orders k of the internal approximation are assumed to be different on various FE of a partition. Orders s j , 0
j m 1 of the boundary approximation are assumed to be the same for all surface
elements of j -layer. Identical partitions of smooth patches of faces of FE into the surface elements are used for different layers, i.e.: r ,l , j
r ,l ,i ,
0 i m 1,
0 j m 1.
In the figures referring to the problems, boundaries of the surface elements are shown in dashes that are perpendicular to the contour of FE. Smooth patches of faces, on which the surface elements are not defined, are marked by dash line placed inside FE.
In Section 3 it was shown that there are two types of the basic functions from the space of conformed by subspace FE:
Functions associated with boundary degrees of freedom of FE. The support of the function is a pair of adjacent FE
Functions associated with internal degrees of freedom of FE. The support of the function is a single FE
Because of this property, the internal degrees of freedom can be computed immediately after the stiffness matrix and right-hand side of FE has been built. So they are not included into the final equation system. In conventional FEM this procedure is called the “condensation” of degrees of freedom. In the tables below N is the dimension of final equation system assembled from the condensed stiffness matrices of FE, so in fact it is the number of boundary degrees of freedom in the system.
Torsion Formulation
Torsion problem consists in finding shear stresses 23 and 13 in cross-section of a bar when it is in twist (Fig. 8.1). This problem can be formulated as a Dirichlet’s problem in two equivalent forms: 1) Find the function w V that meets the equations:
w f
(8.1)
0
(8.2)
0 w
where:
is Laplace’s operator 2 2 i 1 xi 2
0 is the trace operator of zero order
and
V
w
w H 01 ,
w L2
2) Find the function w H 01 that meets the variational equation 2
w v dx x i i
x
For torsion problems f
i 1
fvdx
v H 01
(8.3)
2 in the right-hand side of (8.1) and the function w is called a stress
function. Shear stresses in the section of the bar are determined by the formulas 13
G
w , x2
23
G
w , x1
where is a twist and G is a shear modulus of the material of the bar.
In the external approximation of the equation (8.3) space Gh of the boundary functions consists of only one layer of surface elements.
Model problem Let us consider a model problem when the domain is a square 2 2 and f
1 . The exact
solution of the given problem is known [47]. Due to the symmetry it is sufficient to analyze 1 8 of the square.
In the first experiment the convergence on a family of partitions is under consideration. In Fig. 8.2 are shown the partitions of the domain into triangular FE. For all the partitions and any FE,
the order of the internal approximation is k 1 and the order of the boundary approximation is s 0 . Let us note that for the given problem the external error of approximation is caused not
only by the discontinuity of the approximating functions at the inter-element boundaries, but also by approximate fulfillment of the essential boundary condition (8.2).
In Table 8.1 is given the comparison of the approximate and the exact solutions for the partitions shown in Fig. 8.2. Analyzing the data in Table 8.1 one can see that the convergence rate is not high and in general corresponds to the theoretically predicted order O h . When partitioning
1 8 of the square into 16 FE, the maximum deviation of the function w from zero on the boundary is approximately 6.6 % of its maximum value in the center of the square. For the alternative partitions 5 and 6 the results almost coincide. The results illustrate the convergence of h -version of MEFEA when the mesh of FE is refined without changing the type of FE.
Herewith, both the external and internal errors of approximation simultaneously decrease.
In the second experiment the convergence on a fixed partition of the domain is analyzed. Herewith, the problem is solved on the partition 2(see Fig. 8.2) for different orders s
0;1; 2 of
the boundary approximation, while the order of the internal approximation for each FE of the partition is kept unchanged ( k 4 for all values s ). In this experiment the order of the internal error of approximation is fixed, but the external error is decreased through the refinement of the boundary approximation.
The results given in Table 8.2 show how approximating qualities of the spaces of the boundary functions affect the accuracy of the solution both within and on the boundary of the domain.
When s 2 , the maximum deviation of the function w from zero on the boundary is 1.8% of the function's value in the center of the square. At the transition from s 1 to s 2 , the error at the center of the square is decreased by 15.5 times while at the transition from s 1 to s 2 it is decreased by 1.6 times only. This indicates that when s 0 , due to the low order of the boundary approximation, the major contributor to the total error is the external error. When s 1 both the external and the internal errors become commensurable so that further refinement
of the boundary approximation does not lead to a valuable improvement of the solution inside the domain but it significantly decreases the error in the boundary condition.
Comparison of the Tables 8.1 and 8.2 shows that, for a given number of equations in the final system, the accuracy of MEFEA, for high orders of the internal and the boundary approximations, exceeds the accuracy when using the simplest FE with linear approximating functions.
Stress concentration at a semi-round notch The ratio of the diameter of the bar to the radius of the notch is 2 R r 10 (see Fig. 8.3). Due to the symmetry one can consider only a half of the cross-section by assigning to the symmetry axis the following boundary condition
w 0. n
(8.4)
The domain is simulated with two FE; the contour of the domain is approximated with inscribed broken line. The boundary between the elements is a line that is inscribed into the circumference
with radius 2r . Unlike the previous problem, the orders k of the internal approximation are varied, but the order of the boundary approximation s 0 is unchanged.
For the problem, only smooth patches of the inter-element boundary have boundary degrees of freedom that are not equal to zero. For the smooth patches that coincide with the approximated boundary of the domain, the values of the boundary degrees of freedom are equal to zero due to the homogeneous boundary condition (8.2). It is not necessary to create the boundary degrees of freedom on faces where natural boundary condition (8.4) is prescribed. These patches are marked in Fig. 8.3 with a dash line in accordance with the accepted rule. The number of the equations in the final system is equal to 5; it is defined by the number of the inter-element patches. The minimum order k of the internal approximation that provides P -multi-solvency of the boundary degrees of freedom of FE of the partition equals to 4.
In Table 8.3 are given the errors in shear stresses calculated by the formula 0 0
100% ,
where is the calculated shear stress and 0 is the exact value from [48].
Table 8.3 shows that at an increase of the orders of the internal approximation on both FE from k 4 to k 6 , the error in stress in the area of its concentration (point A in Fig. 8.3) decreases
from 57% to 2.7%. The refinement of the solution is achieved through the increase of the number of the internal degrees of freedom of FE, while the number of the boundary degrees of freedom remains unchanged. However, to get better solution there is no need for a simultaneous increase
of the order of the internal approximation on all FE of the partition. It is sufficient to do that only for those FE located in the area of the stress concentration. Thus, if the order of the internal approximation on the first FE is equal to 6, then the high accuracy is achieved independently of the order on the second FE (see Table 8.3). It happens because the high stress gradient is located in the domain of the first FE, and on the second FE the stresses field is smooth.
It is also necessary to note that the convergence in stresses is the convergence of the first derivatives of the approximated function w . A high accuracy of MEFEA is achieved despite a rather rough partition of the domain into FE, as well as the approximation of the boundary of domain and of the boundary conditions.
Thin plates bending Formulation A great attention is paid to the development of FE for the problem of thin plates bending. It is sufficient to mention a large catalogue [49] of FE used in various software systems (88 elements). The construction of FE for plate bending is more difficult because the differential equation that describes the bending has the fourth order. Bilinear form of the corresponding variational equation contains derivatives of the second order and therefore, the internal approximations of the problem require inter-element continuity not only of the approximating functions but of their first derivatives as well. This results in complex structures of nodal degrees
of freedom of FE, and in high orders of approximating polynomials even for the elements of the simplest form (see, for example the element of Irons in [10]).
The equation of bending of a thin plate of constant thickness is:
2w
q D
(8.5)
where:
2 is a bi-harmonic operator
4 4 4 4 2 2 2 4 x1 x1 x2 x2 2
w is the function of deflection of the plate
q is the intensity of the normal load
D is the cylindrical stiffness of the plate:
D
EH 3 12 1 2 ,
here E is modulus of elasticity; is Poisson's ratio; H is the thickness of the plate
Boundary conditions can be of different types. In case when on the patch 0 of the plate boundary
are given the clamping conditions: w
0
0,
w 0, n
(8.6)
0
and on the patch
1 a lateral load of intensity
Q is applied, then solution of the equation
(8.5) is equivalent to finding the function w W that satisfies the variational equation:
v W
2 w 2v 2 w 2 v 2 w 2v D wv 1 2 x1x2 x1x2 x12 x22 x22 x12 dx qvdx Qvd
where W
1
H 2 is a sub-space of functions that satisfy the cond itions (8.6).
The bending moments and the torque are calculated correspondingly by the formulas:
2 w 2w M11 D 2 2 x2 x1 2 w 2w M 22 D 2 2 x1 x2 M12
2 w D 1 x1x2
For bending problems m 2 and it is necessary to define two layers of surface elements on FE boundary.
Bending of a cantilever plate
Long cantilever plate is loaded on the edge with uniform lateral load intensity Q (Fig. 8.4.a). The input data are as follows: l 10 ; b 2 ; H ratio
1 ; Q 1 ;elasticity modulus E 1 ;Poisson's
0.3 . The dimensions of the plate allow to consider it like a beam, and so to take the
solution of Beam Theory as a target solution to compare with.
The domain is partitioned into two FE (see Fig. 8.4.b), and the following approximation parameters are accepted:
the orders of the internal approximation on each FE: k 4
the orders of the boundary approximation: s j 1, 0 j 1 (i.e. the spaces of polynomials of the first order are the spaces of boundary functions for both layers of surface elements)
The results of MEFEA presented in Fig. 8.4.b practically coincide with the solution of the Beam Theory. It is expected, since the hypothesis of flat section in the Beam Theory is a special case of the hypothesis of straight normal in Kirchhoff’s Thin Plate Theory. Despite discontinuity of approximating functions there are no jumps of deflections and bending moments on the boundary between the elements.
Bending of a square plate It is interesting to know how the complexity of FE shape affects the accuracy. The necessity to use FE of a complicated shape can be caused by the geometry of the initial problem as well as by the need to change the geometry of the domain at iterative computations (for example, when non-linear problems are under consideration). In Section 8.1 it was demonstrated that the use of FE of complex shape provides a high accuracy. However, the problem of torsion is the boundary value problem of the second order. The main purpose of the following experiment is to study the influence of the FE shape on the accuracy of solution of boundary value problems of the fourth order.
Let us consider the following benchmark problem of bending of the square plate: dimensions
2 2 ;thickness H 1 ;loaded by the pressure of intensity q 1 ;modulus of elasticity E 1 ;Poisson's ratio
w
0.3 . On the contour of the plate is assigned either hinge support condition
0 , or clamping conditions (8.6).
Due to the symmetry only 1 / 8 of the plate is simulated. The symmetry condition is that the derivative of the deflection function along the normal to the axis of symmetry being equal to zero. There are three partitions of the domain organized by the degree of their irregularity (Fig. 8.5):
The first partition includes one triangular and two rectangular FE
The second partition is created by the distortion of the first partition through displacement of the common point of FE to the centre of the plate. In result, one FE of swallow-tail type is created
The third partition contains three FE of the swallow-tail type and one FE of irregular hexagon
For all the partitions and for any FE, the order of internal approximation is k 5 . The orders of the boundary approximation s j , 0 j 1 are varied assuming that s0
s1 s . When s 1 ,
for hexagonal FE of the third partition, the order of internal approximation is increased to k 6 in order to provide P -multi-solvency of boundary degrees of freedom of the FE.
Table 8.4 details the results of comparison of MEFEA solution with the exact solution from [50] for the clamped plate problem. One can see that the degree of irregularity of the partition insignificantly affects the accuracy. For all the partitions, when s 0 , the error in deflection at the center of the plate is within the range 30 – 60%; when s 1 the error decreases to 0.1 – 0.2%. So the order of the boundary approximation mainly affects the accuracy: increasing s from s 0 to s 1 , the error in deflections in the center of the plate decreases by two orders of magnitude.
Similar results take place in the case of hinge supports of plate edges (Table 8.5). The deviation from zero of the bending moment M 22 on the edge with the hinge is approximately 2.5% of the maximum value of the moments in the center of the plate.
The magnitude of the discontinuity in moments at inter-element boundaries significantly depends on the order of the boundary approximation. For s 1 it is 0.5 – 3% of the maximum moment. When s 1 , deflection w is calculated with high accuracy for both types of constraints. Its jump on the inter-element boundary is 0.1 – 0.5% of the maximum deflection. The degree of irregularity of the partition practically does not affect the discontinuity.
2D elasticity problems Formulation n Equations that describe equilibrium inside an elastic bod y R have the form:
u u grad div u f
(8.7)
where: n
u
ui i 1 is the displacement vector
f
f i i 1 is the volume force vector
0, 0 are elasticity constants expressed through modulus of elasticity
n
E and
Poisson's ratio as follows:
E
1 1 2
,
E
2 1
is the differential operator of the Theory of Elasticity
Deformations and stresses are expressed through the components of the displacement vector as follows: ij u
u j 1 ui xi 2 x j
,
n ij u ji u kk u ij 2 ij u , k 1 1 i n, 1 j n where: ij
1, i j 0, i j
is Kronecker’s delta.
Let us consider a boundary value problem with the following boundary conditions: u
1
0
(8.8)
n gi , 1 i n u n ij j j 1
(8.9)
2
where:
1 is an open sub-domain of the boundary
1 2
n j , 1 j
n are cosines of the outward normal to the boundary
Let us define the following spaces: V
H 1
Y
L2
V
v v
W
n
n
v
v
V , v Y
V , v 0 1
The problem of finding the function u V that satisfies the relations (8.7) - (8.9) is equivalent to the following variational problem: find the function u
v W
n
n
ij u ij v dx
ij 1
W such that
fi vi dx i 1
n
g v d , i i
(8.10)
2 i 1
where n
f
Y,
12 g H 2 .
The external approximations of the equation (8.10) are constructed on vector FE described in the Section 7.2.3.
Extension of a square plate
The load of intensity q is uniformly distributed along the edges of the plate (Fig. 8.6). The exact values of stresses are 11
12 0, 22 q . Taking into account the symmetry, only 1 4 of
the plate is analyzed at different orders of the internal k and boundary s approximations. The exact solution of the problem is obtained for all combinations of k and s changing in the range
1 k 5, 1 s 3 . Herewith displacements and stresses are continuous at the inter-element boundaries. Therefore, in the example, the external and the internal errors of approximation are equal to zero for all the combinations of the approximation parameters.
Bending of a circular bar
The input data: cross-section of the bar H ratio
H 0.01 0.01 ;radius R 0.2 ; E 0.1 ;Poisson’s
0 . One end of the bar is loaded with total force P 105 , and the other end is
constrained (Fig. 8.7.a).
The curved contour of the bar is approximated by two inscribed broken lines with 16 straight sections each. Three partitions of the domain into FE are considered:
the first partition contains 8 FE, each is a fragment of arch (Fig. 8.7.b)
the second and third partitions are constructed by dividing each FE of the first partition into 2 and 4 FE correspondingly
The order of the boundary approximation is fixed s 1 , and the order of the internal approximation is variable: k 2,3 . The bandwidth of the equation system for all the partitions is M 7 .
MEFEA solution is compared to the solution of the Theory of Circular Bars in [50]. The results of the comparison are summarized in Table 8.6. One can see that, for the first partition at k 2 the error is significant (up to 80% in stress). It is expected, because stresses in the arched FE with 22.5o span (see Fig. 8.7.b) cannot be satisfactory simulated by the polynomial of the first order. Increasing the order of the internal approximation to k 3 , the error in both the displacements and stresses decreases to 10 – 15%. Decreasing the size of FE, the error decreases and for the third partition at k 2 it does not exceed 1%.
One can see that the accuracy is improved either through an increase in the number of FE (i.e. decrease of the internal error) or through an increase in the order of internal approximation. The bar is simulated by one layer of elements only. So the increase in the number of FE at fixed s does not reduce the external error since the size of the inter-element boundaries remains constant. The results indicate that for s 1 the external error, which is caused by the displacement discontinuity on the inter-element boundaries, is small. The main contribution into the error is made by the internal error. This is supported by the fact that displacements are practically continuous. The discontinuity in stresses does not exceed 0.1% of their average value at the discontinuity point.
Extension of a plate with a round hole
The exact solution of the problem of stress concentration around the hole in the infinite plate is known [48]. In computations, the infinite plate is replaced with a square plate with the ratio between the side length and the diameter of the hole a d 8 (Fig. 8.8).
If the classic FEM is applied to a stress concentration problem, then the width of the surrounding the hole layer of elements must be 0.05 – 0.1 of the radius of the hole [51, 52]. Let us use the rather rough partition of 1 4 of the plate (see Fig. 8.8). In this partition, the width of the layer of the adjacent to the hole FE is 1.25 of the radius of the hole. As it follows from the analytical solution of the problem, the area of stress concentration spreads approximately onto just a half of the width of the layer of FE adjacent to the hole.
At a fixed partition of the domain into FE, the number of equations of the system is defined only by the order s of the boundary approximation, and for s 0 the system contains 60 equations while for s 1 it contains 120 equations. Bandwidth of the system at s 0 is M 33 while at s 1 the bandwidth is M 55 .
The results for the stress concentration factor are detailed in Table 8.7 for various combinations of the approximation parameters. Increasing the orders of the internal approximation for the adjacent to the hole FE without increasing the order of the boundary approximation leads to the increase in discontinuities of displacements on the inter-element boundaries (especially in the region, where the stress reaches its maximum). Because of this, the error in stress does not
decrease (see Table 8.7). Instead, at the transition from s 0 to s 1 the error significantly decreases to 1.4%.To increase the accuracy in stress at a fixed partition of the domain it is necessary to increase the orders of both internal and boundary approximations, which agrees with the theoretical results of Section 6.
Special elements in stress concentration This Section contains the results of numerical experiments with the special element developed in the Section 7.2.5 for the simulation of stress concentration regions.
Extension of a plate with an elliptic hole Orientation of long semi-axis of the ellipse, relatively to the direction of the extension, is defined by the angle , and the ratio of the semi-axes is a b
Square plate has ratio 2a L
2 (Fig. 8.9.a).
0.032 between the long axis and the side of the plate. The plate is
partitioned into 9 FE, one of which is a regular octagon with an elliptic hole (Fig. 8.9.b). At one side of the plate is applied a uniformly distributed load while the other side is constrained against vertical displacements.
The order of internal approximation on the complex FE with the hole is k 5 . For the other polynomial FE the order is k 3 . The order of boundary approximation is s 0 . The equation
system contains 33 equations, and the bandwidth of the system is M 23 . The number of Gauss's integration points on faces of the complex FE is N g 8 .
Computations are performed for
30 , 60 , 90 . Orientation of the hole changes without the
change in the partition of the domain. For all the difference between the computed contour stress and the exact stress from the analytical solution for the infinite plate [48] does not exceed 1 – 2%.
Extension of a strip weakened by a square hole with filleted corners
The input data for the problem as shown in Fig. 8.10, are: b d 0.4 ; R b 0.125 ; L d 3 ; d 0.2 ; the thickness of the strip is 0.006;modulus of elasticity is E 19.6 104 ;Poisson's ratio is 0.3 ;load of intensity q 1 is uniformly distributed along the edge.
Taking into account the symmetry, 1/4 of the strip is analyzed by partitioning it into 13 FE (Fig. 8.10.b). The layer of the adjacent to the hole elements, which width is b 8 , includes:
two polynomial elements which simulate straight patches of the contour of the hole
complex element that simulates the fillet
Let us denote through n, k the orders of the internal approximation for the complex and polynomial FE respectively. Values k are assumed to be the same for all polynomial FE.
The results for contour stress, for two versions of the approximation, are shown in Fig. 8.11 where, for comparison purpose, is also given the experimental data from [56]. From Fig. 8.11 follows that the result at n
4, k 5, s 1 almost coincides with the experimental data.
Despite the fact that only one FE simulates straight patches of the contour of the hole, the calculation proves the experimentally observed fact of the stress decrease in the center of the vertical patch of the contour ( x1
b 2 , x2 0 ). The results nicely agree with the data from
work [56], in which 1/4 of the strip is simulated by one complex FE and 81 iso-parametric FE. Decreasing the orders of approximation to n
3, k 4, s 0 , the qualitative picture of the
distribution of stresses on contour remains, but the peak stress decreases approximately by 15% . The effect of the stress decrease in the center of the vertical sides of the contour is not observed as well. Therefore, the accuracy at n
3, k 4, s 0 is insufficient for finding detailed
distribution of stresses in the concentration area.
At n
3, k 4, s 0 the size of the equation system is N 66 and its bandwidth is M 30
. At n
4, k 5, s 1 one has N 132 and M 61 .
Extension of a perforated plate
The perforation of the plate corresponds to a riveted or bolted joint (Fig. 8.12.a). The input data: the diameter of the holes d 0.02 ; the intensity of the load 1 ; modulus of elasticity
E 19.6 19.6 104 ;Poisson's ratio
0.3 .
Due to the symmetry, 1/4 of the plate is analyzed. The domain is partitioned into 5 complex FE with the holes and 12 polynomial FE (Fig. 8.12.b). The order of the internal approximation is n 5 for the complex FE and k 4 for the polynomial FE, the order of the boundary
approximation is s 1 .
The stress distribution at contours of the holes is shown in Fig. 8.12.c. From Fig. 8.12.c it it follows that the mutual influence of the holes, that are in the central area of the plate, causes some shift of the stress peak towards the neighbor hole. A similar effect is observed for an infinite plate with two rows of holes [57]. However, assuming K max 4.85 the stress concentration factor for the infinite plate, the actual value for the finite plate is in the range 5.40 – 5.85. For the holes close to the edge of the plate, the shifts of the stress peaks do not occur because of the influence of the edge.
For the problem, the order of the equation system is N 130 , and the bandwidth is M 51 .
Strip with a round hole The input data: the ratio of the radius of the hole to the width of the strip is R b 0.275 (Fig.8.13.a);Poisson's ratio is
0.3 ;one edge of the strip is fixed: u1 u2 0 at x1 0 , and
at the opposite edge the load is applied (Fig.8.13.b).
The following load cases are analyzed:
extension (Fig.8.13.b)
lateral bending (Fig. 8.13.c) pure bending (Fig.8.13.d)
The partition of the strip includes three FE: one FE is complex (element with the hole), and the other two are polynomial (see Fig. 8.13.b). The number of Gauss’s integration points on the boundary patches of the complex FE is N g 9 .
Stress concentration factor is determined by formula K
A Bnom
,
nom where: A is the normal stress at point A (see Fig. 8.13.b) calculated by MEFEA; B is the
nominal stress at point B calculated using the elementary formulas of the Theory of Strength of Materials for the section weakened by the hole.
In Table 8.8 are given the stress results for the extension problem depending on the orders of internal and boundary approximations. For comparison purpose, in Table 8.8 are also given the results obtained by classic FEM with 8-node iso-parametric elements (MIFE), by the method of large elements (MLE), as well as experimental data of p hoto-elasticity method (MPE) [54].
From Table 8.8 it follows that the best conformity of the results of MEFEA with the results of other methods takes place when s 1 . Herewith, the number of elements in MEFEA is 3 times less than in MLE and 90 times less than in MIFE. Similar results take place for the lateral bending (see Table 8.9) and pure bending (see Table 8.10).
Extension of a strip with a cutout
This problem illustrates the effectiveness of the complex FE with an elliptic arch (see Fig. 7.4.b). The input data (see Fig. 8.14.a): the ratio of the depth of the cutout to the width of the strip is
R b 0.25 ; Poisson’s ratio is 0.3 .
Variants of the partitions of the strip are shown in Fig. 8.14.b.c. The data detailed in Table 8.11 indicate that the results of MEFEA are similar to the results of other methods [54]. Herewith the same accuracy is achieved at smaller orders of the equation system.
2D problems of linear Fracture Mechanics Stress intensity factors are the main parameters of Fracture Mechanics to be evaluated. They characterize the stress field at the tip of a crack. They are used to forecast survivability of the structures containing cracks as well as in fatigue analysis. Accurate simulation of the stress field at the tip of the crack is critical to Fracture Mechanics. For a qualitative approximation of the solution in the vicinity of the crack special FE are required that have the singularity in stresses of type r 1 2 [58 - 68]. The problem of construction of such elements is twofold: on one hand, it is necessary to simulate the singularity of a certain type on the element, on the other hand it is necessary to provide the continuity of the approximating functions (including singular) through the boundary of the element.
The simulation of the singularity itself does not pose any major difficulty and it can be done in various ways:
by modifying the conventional technology of MFE (for example, by node shifting in iso-parametric FE [66, 69 - 71])
by employment of auxiliary singular functions [9, 72 - 75]
by use of the local asymptotic representations of solutions for the elements with singularities [76 - 78]
More difficult problem is to provide continuity of the approximating functions through the boundary of singular FE. This problem is solved successfully when using modified iso-parametric FE. However, as it has been demonstrated (see [79 - 81]), the size of the element has a major influence on the accuracy, and it leads to the necessity of a mesh refinement near the tip of the crack. Thus, for example, in work [69] when solving the problem of inclined crack the size of the smallest iso-parametric element was 1 250 of the length of the crack.
Discontinuity of the approximating functions can be handled with special buffer elements that surround the element with singularity [82], or by developing singular asymptotic representations that turns to zero on the boundary [83, 84], or by hybrid schemes of FEM. Hybrid FE are widely used when solving the problems of fracture mechanics. They are based on independent approximation of singular field of stresses and displacements of the boundary of FE [54, 59, 85 90]. Complexity of the construction of conformed singular FE makes some authors use the nonconformed elements in practical calculations without an appropriate theoretical justification [67,
81, 91, 92]. In this case the acceptability of FE is decided through numerical experiments on benchmark problems. However, the benchmark problems cannot cover all the diversity of practical cases and the successful solution of a problem does not guarantee the success in all situations.
Below, the special FE described in the Section 7.2.5 is used to solve fracture mechanics problems. Herewith the length of one of the semi-axes of the elliptic hole is equal to zero, and the space P K of the approximating functions of FE is generated by the terms of the complete asymptotic representation of the exact solution of the problem with a crack in infinite domain. Let us notice that the FE allows the crack orientation change and growth without re-partitioning of the domain. It is convenient when solving the problems of the crack propagation.
A plate with an inclined crack
The exact stress intensity factors for the inclined crack in the infinite plate are calculated by the formulas K I
l sin 2
K II
l sin cos
where: is the stress at the infinity; l is the semi-length of the crack; is the orientation angle (Fig. 8.15.a).
In the computations, the infinite plate is replaced with a square plate with a ratio of the length of the crack to the side of the square of 2l L
0.032 (Fig.8.15.b). Uniformly distributed load
1 is applied at the upper edge, while the opposite edge is constrained in normal direction.
The partition includes one complex FE with the crack and 8 polynomial FE (see Fig. 8.15.b). The input data: l 5 ; E 1 ; 0.3 ;
15 , 30 , 60 , 90 .
The crack orientation is changed without re-partitioning the domain. The orders of the internal approximation are: n 5 for the complex FE, and k 3 for all polynomial FE. The order of the boundary approximation is s 0 . The size of the equation system is N 34 , the bandwidth is M
24 .
In Table 8.12 are given the errors in stress intensity factors K I and K II at the tip of the crack for different orientations. Despite the rather coarse partition of the domain and the low order of the boundary approximation, a high accuracy in fracture parameters is achieved. For small , the error for K I and K II is rather lower than for large . Thus, at K I and 0.8% for K II ; at
15 , the error is 2.8% for
30 the error is 0.5% for K I and 0.2% for K II . Values of K I
and K II , calculated for both tips of the crack, coincide up to three digits. Error in the load boundary conditions of the problem ranges 0.05 – 0.1%.
The analysis of stress field shows that on the contour of the complex FE takes place the one-axial stress state that corresponds to the extension of the plate without the crack. The domain of high stress gradients is located completely inside the complex FE. This is caused by that the ratio
0.125 between the length of the crack and the diameter of the complex FE is rather small. Interelement boundaries are outside the area of stress disturbance and this is the reason of high accuracy at the low order of the boundary approximation.
Strip with a crack
In the case when the length of the crack is of the same order of magnitude as the width of the strip, the stress intensity factors cannot be calculated by the formulas for the crack in the infinite domain. The finite width of the strip is usually accounted by the correlation factor, i.e. the actual intensity factor is calculated by formula K I
FK I
where K I is the intensity factor for the crack in infinite domain.
The data to the problem: the ratio of the width to the length of the strip is a L
5 18 ;module of
elasticity is E 19.6 104 ;Poisson’s ratio is 0.3 ;one edge of the strip is fixed while at the opposite edge the uniformly distributed load is applied (Fig. 8.16).
The strip is simulated by three FE, the element with the crack is complex,the other two elements are polynomial (Fig. 8.16).The order of the internal approximation on the polynomial FE is k 4 ,and on the complex FE n 5 . The order of the boundary approximation is s 1 .
The length of the crack is changed without re-partitioning the domain. From Table 8.13 it follows that the results of calculation of the correlation factor conform to the data from other theories given in work [93].
Axisymmetrical elasticity problems Formulation In polar coordinate system x1 , x2 , x3 , the equilibrium equations (8.7) have the form
11 12 11 33 f 1 0 x1 x2 x1 12 22 12 f 2 0 x1 x2 x1 where stresses are defined by the formulas (see Fig. 8.17):
11
u u2 u1 D 1 x1 1 x2 1 x1
22
u1 u2 u1 D 1 x1 x2 1 x1
33
u1 u u2 D 1 1 x1 1 x2 x1 12
D
D
1 2 u1 u2 x1 2 1 x2
E 1
1 1 2
here: E is the modulus of elasticity; is Poisson's ratio; u1 , u2 are, correspondingly, radial and axial displacements; f1 , f 2 are the radial and axial components of the volume force.
The variational problem (8.10) is transformed as follows: find the function u W that satisfies the equation
3 ii u ii v 12 u 12 v dx i 1 *
2
2
f v dx g v d i i
i i
* i 1
(8.11)
v W
2 i 1
where:
ii v 33
vi , xi
i
1, 2
v1 x1
12 v
v1 v2 x2 x1
W
V1 V 2
V1
u1 H 1
V2
H 1
u1
0 for x1 0
and where is the half of the meridian section of the structure located in the semi-plane x1 0 .
External approximations of the equation (8.11) are built with axisymmetrical elements described in Section 7.2.4.
Hollow sphere under internal pressure
The data to the problem: the radius of the sphere is R H
0.1 ; the thickness of the wall is
0.04 ; the internal pressure is P 1 ; the material properties are E 2 105 , 0.3 .
For this problem displacements and stresses are functions of the distance from the center of the sphere. Therefore, for the analysis of the sphere, one can consider just a fragment of its meridian section that is bounded by the central angle of 15 . The domain is the polygon ACDFB (Fig. 8.18.a) with its lines AC and DFB inscribed into the contour of the meridian section of the sphere. The displacements along the normal to sides AB and CD of the fragment ACDBF equal zero. The domain is partitioned into one and two FE (Fig. 8.18.a.b).
The results detailed in Table 8.14 indicate the high accuracy achieved even at the partition of the domain into one FE. Thus, at s
1, k 4 , the error in maximum radial stress 11 is 1.1% and
in maximum circumference stress 33 is 6.4%. The radial stress at the outer surface where it must equal zero is 3% of the maximum. The error in circumference stresses on the outer surface is 0.05%. Partitioning the domain into two FE (Fig.8.18.b), the accuracy increases.
Stress concentration around a spherical cavity
The infinite domain around the cavity (Fig. 8.19.a) is simulated by a finite cylinder (Fig. 8.19.b). The data to the problem: the radius of the cylinder is R ; the height of the cylinder is 2 R ; the ratio between the radius of the cylinder and the radius of the cavity is R r 8 ;Poisson's ratio is
0.3 ; the uniformly distributed load of unit intensity is applied to the ends of the cylinder
while its side surface is free.
In Fig. 8.19.b is shown the partition of 1/4 of the meridian section of the cylinder. Computations are performed for two partitions of the domain that differ by the width a of the layer of the adjacent to the cavity elements (Fig. 8.19.b). For the first partition a r 1.25 , and for the second one a / r 0.35 .
The order of the boundary approximation is increased from s 0 to s 2 . The order of the internal approximation for all s and for all FE is k 4 .
The results for stresses on the equator and in the pole of the spherical cavity (points A and B in Fig. 8.19.b) are given in Table 8.15 for different partitions and orders of the boundary approximation. One can see that the width of the boundary layer of FE significantly influences the accuracy. Thus, for a r
a/r
0.35, s 2 ,
1.25, s 2 , the error in maximum stress is 17.8% and, for
it is 0.68%. The error rapidly decreases by increasing the order of the
boundary approximation that proves that for k 4 the internal error of approximation is rather small and the accuracy is, in general, determined by the quality of the boundary approximation.
Round plate under pressure
The data for the problem: the radius of the plate is R pressure is q
0.1 ; the thickness is H 0.01 ; the
1 ; the material properties E 106 , 0.3 .
The domain is partitioned into two and four FE (Fig. 8.20). The order of the boundary approximation for both partitions is s 1 . The order of the internal approximation k is increased from 2 to 4.
From Table 8.16 it follows that increase in the order of the internal approximation from k 2 to k 3 provides sufficient accuracy of the solution. The further increase to k 4 practically does
not cause any changes in the displacements and stresses. So the internal error is significant for
s
1, k 2 . For s 1, k 3 the internal and external errors become of the same order of
magnitude and the increase to k 4 does not lead to a valuable change in the accuracy. For k 4 , the difference between the calculated stress and the Theory of Thin Plates Bending is
approximately 0.5% while the difference in the displacements in the center of the plate is 4.1%.
Spherical dome under pressure
The data to the problem: the radius of the middle surface of the dome is R is H
2.28 ; the thickness
0.076 ; the angle is 35 ;Poisson's ratio is 1 6 ;the pressure is q 7 .
The dome is partitioned uniformly with spacing
3.5 (Fig. 8.21). Three variants of the
internal approximation are analyzed: 1. k 2 for all FE 2. k 3 for all FE 3. For the FE located at the edge of the dome k 4 while for the others k 3
The order of the boundary approximation is s 1 .
In Table 8.17 by the asterisk is marked the data that correspond to the third variant of the approximation. Here the stresses on the external and internal surfaces of the dome are marked by (+) and (-) respectively. The data in Table 8.17 show that the results in 3D formulation well agree with the theory of thin shells. The increase of the order of the internal approximation for the FE adjacent to the edge of the dome (third variant of the approximation) does not lead to the change in stresses at the center of the dome. At the edge of the dome the stresses also change insignificantly (approximately by 3%).
Sine-shaped shell of infinite length under pressure
Equations of the middle surface of the shell in the polar coordinate system (Fig. 8.22.a) are as follows: x1
R f 1 cos
x2
l
2
where is a parameter.
Considering the periodic symmetry of the structure, only one semi-wave of the shell is analyzed under the displacement boundary conditions: u2 R
0 at 0, . The other data:
0.225 ; f 0.025 ; l 0.1 ; H 0.0025 ;the material properties E 2 105 , 0.3 ;the
pressure q
1.
For better approximation of the geometry of the shell, the semi-wave of is partitioned into 18 elements (Fig. 8.22.b). The order of the internal approximation is k 2 for all FE,the order of the boundary approximation is s 1 . The number of the equations in the system is N 58 , the bandwidth is M 7 .
In Fig. 8.23 are shown the results for meridian 1 and circumference 2 stresses on external (+) and internal (-) surfaces. For the comparison purposes, in Fig. 8.23 are also given the results obtained by Timoshenko's model [94] as well as by the Theory of Thin Shells [95]. From Fig. 8.23 it follows that the calculations by different models give similar results.
Shell compensator
In Fig. 8.24 is shown a semi-wave of the compensator and its partition into 14 axisymmetric elements. The semi-wave consists of two torus segments I and III , and a circular plate II . Coordinate is measured along the meridian of the compensator. The modulus of elasticity is
E 2.1 105 MPa, the Poisson’s ratio is
0.3 .
The order of the internal approximation for all FE is k 3 , the order of the boundary approximation is s 1 .
Two load cases are analyzed.
Case 1 - To the upper rim of the compensator compressing axial force P
9810 N is applied and
there are no restrictions for the displacements of the rim. The condition of symmetry u2 defined at x2
0 is
0 . In Fig. 8.25 are shown the graphs for the meridian stress 1 and for the
circumference stress 2 , both on the internal (continuous lines) and on the external (dash line) surfaces of the compensator.
From Fig. 8.25 it follows that the maximum circular stress occur on the upper rim of the compensator, and the stresses at the internal and external surfaces are practically identical. Meridian stress has two obvious peaks: the first peak is located in the second half of the torus segment I while the second peak is located at the center of the torus segment III . Herewith, the amplitude of the second peak exceeds the amplitude of the first peak approximately by 50%.
Case 2 - Often the compensator system consists of few connected shell compensators. The calculations are made for the compensator that is one wave of the compensator system, at the given axial convergence of its rims. Boundary conditions of the problem are defined as follows:
0 at x2 0
u2
0.027 m (upper rim) u2 0.001 m at x2
The stress graphs are given in Fig. 8.26. One can see that the stress picture significantly depends on the type of the boundary conditions at the upper rim. For a given convergence of the rims, maximum circumference stress occurs at the tips of the torus segments. At the upper rim the meridian stress is significant. The second peak of the meridian stress is approximately equal to
the amplitude of the first peak and it is located at the center of the torus segment III . The stress graphs shown in Fig. 8.25 well conform to the results for such compensators from [94].
Figure legends for Section 8:
Fig. 8.1 Shear stresses in torsion
Fig. 8.2 Partitions
Fig. 8.3 Partition of the cross-section of the bar
Fig. 8.4 Bending of a cantilever plate: (a) dimensions and loading (b) partition and results
Fig. 8.5 Partitions of 1/8 of the plate
Fig. 8.6 Loading and partition
Fig. 8.7
Bending of a circular bar: (a) geometry and boundary conditions (b) partitions (only 1/8 of the bar is shown)
Fig. 8.8 Dimensions, loading, and partition of 1/4 of the plate
Fig. 8.9 A plate with an elliptical hole: (a) infinite plate (b) finite element model
Fig. 8.10 A strip with a square hole: (a) geometry and loading (b) partition of 1/4 of the plate
Fig. 8.11 Stress at the contour of the hole, where: Broken line is the result for k Solid line is the result for k circles is experimental data
3, n 4, s 0 ;
4, n 4, s 1 ;
Fig. 8.12 A perforated plate: (a) geometry and loading (b) partition (c) Stress distribution at the contour of the holes (circles show the points where the stress reaches its maximum)
Fig. 8.13 A strip with a round hole: (a) geometry (b) partition (c) (d) load cases
Fig. 8.14 A strip with a round cut at its side: (a) geometry (b) (c) partitions
Fig. 8.15 A plate with an inclined crack: (a) infinite plate (b) finite element model
Fig. 8.16 A strip with a crack: geometry, loading, and partition
Fig. 8.17 Stresses in the axisymmetrical problem
Fig. 8.18 A hollow sphere under internal pressure: (a)(b) partitions
Fig. 8.19 Stress concentration around spherical cavity: (a) infinite domain (b) finite element model
Fig. 8.20 Bending of a round plate Fig. 8.21 A spherical dome under external pressure
Fig. 8.22 A sine-shaped shell under internal pressure: (a) geometry
(b) partition of a semi-wave
Fig. 8.23 Stress distribution: Solid line is data from [94], Squares are results from [95], Circles is MEFEA solution
Fig. 8.24 Partition of a semi-wave of a shell compensator
Fig. 8.25 Stresses under axial force
Fig. 8.26 Stresses under axial displacement of the rim
Reference list 1. Baseley G.R., Cheung Y.K., Irons B.M., Zienkiewicz O.C., “Triangular elements in bending-conforming and nonconforming solutions”, Proceedings of the 1 st conference on matrix methods in structural mechanics , Dayton, 1965, pp. 547-576.
2. Amores D., Deserable D., Martinez I., Surry C., “ Elements finis conformes on non conformes en elasticite plane on de revolution”, Stud. geotechn. et mech ., 1981, 3, 1, pp. 33-34. 3. Bergan P.G., Nygard M.K., “Finite elements with increased freedom in choosing shape functions”, Int. J. Numer. Meth. Eng., 1984, 20, 4, pp. 643-663. 4. Fortin M., Soulie M., “A nonconforming piecewise quadratic finite element on triangles”, Int. J. Numer. Meth. Eng., 1983, 19, 4, pp. 505-520.
5. Irons B.M., Razzaque A., “Experiens with the patch test for convergence of finite elements”, Proc. Symp. on the Mathematical Foundations of the Finite Element Method with Applications to Partial Differential Operators , Baltimore; N.Y., 1972, pp. 557-587.
6. Loikkanen M.J., Irons B.M., “An 8-node brick finite element”, Int. J. Numer. Meth. Eng., 1984, 20, 3, pp. 523-528. 7. Zienkiewicz O.C., Qu S., Taylor R.L., Nakazawa S.,“The patch test for mixed formulations”, Int. J. Numer. Meth. Eng., 1986, 23, 10, pp. 1873-1883. 8. Свойский Ф.М.,“Несовместные конечные элементы для расчета листовых пространственных конструкций”, Прикладные проблемы прочности и пластичности, Горький, 1986, № 33, c. 59-70. 9. Gilbert Strang, George J. Fix, An Analysis of the Finite Element Method , Prentice-Hall, 1973, 349 p.
10. Philippe Ciarlet,The Finite Element Method for Elliptic Problems , North-Holland Publishing Company, 1978, 512 p. 11. Roger Temam, Navier-Stokes Equations , North-Holland Publishing Company, 1979, 408 p. 12. Arantes e Oliveira E.R ., “ The patch test and general convergence criteria of the finite element method”, Int. I. of Solids and Structures, 1977, 13, pp. 159-178. 13. Janovsky V., Prochazka P., “Convergence analysis of a nonconforming finite element method solving a plate with ribs”, Apl. mat ., 1978, 23, 1, pp. 9-30. 14. Janovsky V., Prochazka P ,.” The nonconforming finite element method in the problem of clamped plate with ribs”, Apl. mat., 1976, 21, 4, pp. 273-289. 15. Lascaux P., Lesaint P., “Some nonconforming finite elements for the plate bending problem”, Rev. franc. automat., inform., rech. oper. , 1975, 9, R1, pp. 9-53. 16. Lesaint P., Zlamal M., “Convergence of the nonconforming Wilson element for arbitrary quadrilateral meshes”, Numer. Math., 1980, 36, 1, pp. 33-52. 17. Nitsche J.A., “Convergence of nonconforming methods”, Math. Aspects Finite Elements Partial Differ. Equat. , 1974, pp. 15-53.
18. Rannacher R., “L - Feherabschatzung fur eine nicht Ronforme Finite-Elemente-Methode beim Plattenproblem”, Z. agnew. Math. und Mech., 1977, 57, 5, pp. T247-T249. 19. Rannacher R. , “ Nonconforming finite element methods for eigenvalue problems in linear plate theory”, Numer. Math., 1979, 33, 1, pp. 23-42. 20. Rannacher R ., “On nonconforming and mixed finite element method s for plate bending problems. The linear case”, RAIRO. Anal. Numer ., 1979, 13, 4, pp. 369-387.
21. Stummel F., “Basic compactness properties of nonconforming and h ybrid finite element spaces”, RAIRO Anal. numer ., 1980, 14, 1, pp. 81-115. 22. Stummel F., “The generalized patch test”, SIAM J. Numer. Anal ., 1979, 16, 3, pp. 449471. 23. Stummel F., “The limitations of the patch test”, Int. J. Numer. Meth. Eng., 1980, 15, pp. 177-188. 24. Евзеров И. Д.,“Достаточные условия сходимости метода конечных элементов при использовании несовместных аппроксимирующих функций”,Сопротивление материалов и теория сооружений, Киев, 1980, № 36, c. 127 – 130. 25. Евзеров И. Д.,“Неконформные конечные элементы в задаче на собственные значения», Численные методы механики сплошной среды, 1984, 15, 5, c. 84 – 90. 26. Евзеров И. Д.,“Оценки погрешности по перемещениям при использовании несовместных конечных элементов»,Численные методы механики сплошной среды, 1983, 14, 5, c. 24 – 31. 27. Корнеев В.Г., Схемы метода конечных элементов высоких порядков точности, Л., 1977, 206 с. 28. Лапин А.В, Сеточные аппроксимации вариационных неравенств, Казань, 1984, 96 с. 29. Barnhill R.E., Brown J.H., “Nonconforming finite elements for curved regions”, Lect. Notes Math., 1976, 506, pp. 1-14.
30. Goldstein C.I., “Variational crimes and Lerror estimates in the finite element method”, Math. Comput ., 1980, 35, 152, pp. 1131-1157.
31. Samuelsson A., “The global constant strain condition and the patch test”, Energy Methods Finite Elem. Anal. , 1979, pp. 47-58.
32. Ce’a., “Approximation variationelle des problemes aux limites”, Ann. Inst. Fourier ., 1964, 14, pp. 345-444. 33. Jean-Pierre Aubin, Approximation of Elliptic Boundary-Value Problems , JohnWiley&Sons, 1972, 383 p. 34. Glowinski R., “Approximations extrenes, par elements finis de Lagrange d’ordre un et deux, du probleme de Dirichlet pour l’operateur biharmonique. Method iterative de resolution d es problemes approaches”, Top. Numer. Anal ., L.;N.Y., 1973, pp. 123-171. 35. Филин А.П., Коносов Г.И.,“Комбинированное использоание идей метода конечных элементов и классических краевых методов при решении краевых задач»,Устойчивость и динамика судовых конструкций, Л., 1985, c. 103– 112. 36. Филин А.П., Коносов Г.И.,“Решение краевой задачи при новом ( смешанном ) методе построения базиса», Строительная механика и прчность судовых конструкций, Л., 1985, c. 103– 112. 37. Li Zi-Cai, ”On the reduced rate of convergence for a nonconforming combined method”, SIAM J. Numer. Anal. , 1983, 20, 1, pp. 86-93.
38. Li Zi-Cai, ”A nonconforming combined method for solving Laplace’s boundary value problems with singularities”, Numer. Math., 1986, 49, 5, pp. 475-497. 39. Li Zi-Cai, “An approach f or combining the Ritz-Galerkin and finite element methods”, J. Approxim. Theory. , 1983, 39, 2, pp. 132-152.
40. Литвинов В.Г.,“Об одной модификации метода Ритца для вариационных уравнений и ее приложении к краевым задачам со смешанными граничными условиями», Дифференциальные уравнения, 1981, 17, 3, c. 519 – 526. 41. Blair J.J., “Higher order approximations to the boundary conditions for the finite element method”, Math. Comput ., 1976, 30, 134, pp. 250-262. 42. Квек В., Корнеев В.Г.,“О точности вычислительных схем с несовместными конечными элементами Вильсона”, Вычислительная и прикладная математика, Киев, 1983, 50, c. 17-23. 43. Kikuchi F., “Convergence of the ACM finite element scheme for plate bending problems”, Publs Res. Inst. Math. Sci. , 1975, 11, pp. 247-265. 44. Miyoshi T., “Convergence of finite element solution, represented by a non-conforming basis”, Kumamoto J. Sci. (Math). , 1972, 9, 1, pp. 11-20. 45. Shi Zhong-ci., „A convergence condition for the quadrilateral Wilson element”, Numer. Math., 1984, 44, 3, pp. 349-361.
46. Shi Zhong-ci., “A remark on the optimal order of convergence of Wilson’s nonconforming element”, Math. numer. Sin., 1986, 8, 2, pp. 159-163. 47. Михлин С.Г., Смольский Х.Л., Приближенные методы решения дифференциальных и интегральных уравнений, М., 1965, 384 с. 48. Демидов С.П.,Теория упругости, М., 1979, 432 с. 49. Hrabok M.M., Hrudey T.M., “A review and catalogue of plate bending finite elements”, Comput. and Struct. , 1984, 19, 3, pp. 479-495.
50. Прочность. Устойчивость. Колебания. Справочник, В 3 т. , М., 1968, 1, 831 с.
51. Janovsky V., Prochazka P., “Convergence analysis of a nonconforming finite element method solving a plate with ribs”, Apl. mat., 1978, 23, 1, pp. 9-30. 52. Сахаров А.С. и др. Метод конечных элементов в механике твердых тел, Киев, Лейпциг, 1982, 479 с. 53. Delves L.M., Hall C.A., “An implicit matching principle for global elemen t calculations”, J. Inst. Math. and Appl., 1979, 23, 2, pp. 223-234.
54. Jirousek J., “Implementation of local effects into conventional and nonconventional finite element formulations”, Local Eff. Anal. Struct., 1985, pp. 279-298. 55. Jirousek J., Teodorescu P., “Large finite elements method for the solution of problems in the theory of elasticity”, Comput. and Struct ., 1982, 15, 5, pp. 575-587. 56. Moriya K., “Finite element analysis of stress concentration around elliptic holes, fillets and cracks in plates under stretching and out-of- plane bending”, Bull. JSME., 1984, 27, 234, pp. 2611-2618. 57. R.E.Peterson,Stress Concentration Factors , John-Wiley&Sons, 1974, 300p. 58. Chen E.P., “Finite element analysis of a bimaterial interface crack”, Theor. and Appl. Fract. Mech., 1985, 3, 3, pp. 257-262.
59. Kuna Meinhard, “Behandlung raumlicher Ribprobleme mit d er Methode der finiten Elemente”, Techn. Mech. , 1984, 5, 3, ff. 23-25. 60. Lin K.Y., Mar J. W., “Finite element analysis of stress intensity factors for cracks at a bimaterial interface”, Int. J. Fract., 1976, 12, 4, pp. 521-531. 61. Ogen G., Schiff B., “Constrained finite elements for singular boundary value problems”, J. Comput. Phys., 1983, 51, 1, pp. 65-82.
62. Souchet Rene, “Solution of two-dimensional elastic crack problems using a localized finite element method”, Eng. Fract. Mech., 1984, 20, 1, pp. 169-177. 63. Wahba N.N., “On the use of singular displacement finite elements for cracked plate in bending”, Int. J. Fract., 1985, 27, 1, pp. 3-30. 64. Wait R., “Singular isoparametric finite elements”, J. Inst. Math. and Appl., 1977, 20, 1, pp. 133-141. 65. Ye T.Q., Gallagher R.H., “A singular finite element for analysis in plate bending problem in fracture mechanics”, Int. J. Fract., 1984, 24, 2, pp. 137-147. 66. Гришин В.И., Бичеев Т.К.,“Применение сингулярных эле ментов при решении двумерных и трехмерных задач механики разрушени», Проблемы прочности, 1986, 12, c.91-96 67. Дашевский Е.М., Роттер М.В.,“Применение метода суперэлемента к решению задач механики разрушения с подрастающей трещиной”, Проблемы точности, 1984, 7, c.12-17. 68. Игнатюк Н.Н., Никишков Г.П.,“Сингулярный конечный элемент с произвольной степенью асимптотики”, Исследование прочности материалов и конструкций атомной теники, М., 1984, c. 8 - 12. 69. Морозов Е.М., Никишков Г.П. Метод конечных элементов в механике разрушения, М., 1980, 256 с. 70. Saouma Victor E., Schwemmer David, “Numerical evaluation of the quarterpoint crack tip element”, Int. J. Numer. Meth. Eng., 1984, 20, 9, pp. 1629-1641. 71. Staab G.H., “Estimating stress intensity factors with singular components of the total finite element solution”, Int. J. Numer. Meth. Eng., 1982, 18, 7, pp. 1063-1076.
72. Марчук Г.И., Агошков В.И. Ввведение в проекционно-сеточные методы, М., 1981, 216 с. 73. Blum H., Dobrowolski M., “On finite element methods for elliptic equations on domains with corners”, Computing. , 1982, 28, 1, pp. 53-63. 74. Destuynder P., Djaona M., “Estimation de l’erreur sur le coefficient de la singularite de la solution d’un probleme elliptique sur un ouvert avec coin”, RAIRO. Anal. numer., 1980, 14, 3, pp. 239-248.
75. Yamamoto Y., Tokuda N., Sumi Y.,”Finite element treatment of singularities of boundary value problems and its application to analysis of stress intensity factors”, Theory and Pract. Finite Elem. Struct. Anal. , Tokyo, 1973, pp. 75-90.
76. Hendry J., “Singular problems and the global element method”, Comput. Meth. Appl. Mech. and Eng., 1980, 21, 1, pp. 1-15.
77. Kermode M., Mc Kerrel A., Delves L.M., “The calculation of singular coefficients”, Comput. Meth. Appl. Mech. and Eng. , 1985, 50, 3, pp. 205-215.
78. Tolley M.D., “Torsion des barres polygonales”, Bull. cl. sci. Acad. rog. Belg., 1977, 63, 11, pp. 902-912. 79. Harrop L.P., “The optimum size of quarter - point crack TIP elements”, Int. J. Numer. Meth. Eng., 1982, 18, 7, pp. 1101-1103.
80. Jehia Nabil A.B., Shephard Mark S., “On effect of quarter -point element size on fructure criteria”, Int. J. Numer. Meth. Eng., 1985, 21, 1, pp. 1911-1924. 81. Yunus S.M., Shephard M.S., “A comment on the effectiveness of uniform finite element refinements”, Int. J. Numer. Math. Eng., 1984, 20, 1, pp. 187-194.
82. Лущик О.Н.,“Модификация метода конечных элементов для областей с особыми точками”, Изв. АН СССР. Сер. МТТ, 1986, 4, c. 87 – 93. 83. Li Yao-Chen, “The finite element method by employing the singular element with concordant displacement at the crack tip”, Eng. Fract. Mech., 1984, 19, 5, pp. 959-972. 84. Stern Morris, Becker Eric B., “A nonconforming crack tip element with quadratic variation in the singular fields”, Int. J. Numer. Meth. Eng., 1978, 12, 2, pp. 279-288. 85. Chen Wen-Hwa, Chen Pei-Yen, “A hybrid-displacement finite element model for the bending analysis of thin cracked plates”, Int. J. Fract., 1984, 24, 2, pp. 83-106. 86. Orringer O., “ K -solutions with assumed-stress hybrid elements”, J. of the Structural Division, Proc. Of the Amer. Soc. Of Civil Engin. , 1977, 103, 2, pp. 321-334.
87. Pian T. H. H., “Derivation of Elements Stiffness Matrices by Assumed Stress Distributions”, American Institute of Aeronautics and Astronautics Journal, 1964, 2, July, pp. 1333-1336. 88. Piltner R., “Special finite elements for an appropriate treatment of local e ffects”, Local Eff. Anl. Struct., 1985, pp. 299-314.
89. Piltner R., “Special finite elements with holes and internal cracks”, Int. J. Numer. Meth. Eng., 1985, 21, 8, pp. 1471-1485.
90. Wang S.S., Yuan F.G., “A hybrid finite element approach to composite laminate elasticity problems with singularities”, Trans. ASME : J. Appl. Mech. , 1983, 50, 4a, pp. 835-844. 91. Byskov Esben, “The calculation of stress intensity factors using the finite element method with cracked elements”, Int. J. Fract., 1984, 26, 4, pp. 329-337.
